{"id":100814,"date":"2026-09-05T12:33:50","date_gmt":"2026-09-05T04:33:50","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/100814.html"},"modified":"2026-09-05T12:33:50","modified_gmt":"2026-09-05T04:33:50","slug":"%e8%bf%81%e7%a7%bb%e5%ad%a6%e4%b9%a0%e6%80%8e%e4%b9%88%e8%90%bd%e5%9c%b0%ef%bc%9ftransformers-%e5%ba%93%e5%be%ae%e8%b0%83%e5%ae%9e%e6%88%98","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/100814.html","title":{"rendered":"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730\uff1fTransformers \u5e93\u5fae\u8c03\u5b9e\u6218"},"content":{"rendered":"<h2>\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730&#xff1f;Transformers \u5e93\u5fae\u8c03\u5b9e\u6218<\/h2>\n<p>\u5173\u952e\u8bcd&#xff1a;\u8fc1\u79fb\u5b66\u4e60\u3001Transfer Learning\u3001Hugging Face Transformers\u3001\u6a21\u578b\u5fae\u8c03&#xff08;fine-tuning&#xff09;\u3001\u9884\u8bad\u7ec3\u6a21\u578b<\/p>\n<p>\u9002\u8bfb\u4eba\u7fa4&#xff1a;\u624b\u91cc\u6709\u51e0\u767e\u5230\u51e0\u4e07\u6761\u6807\u6ce8\u6570\u636e\u3001\u60f3\u8bad\u4e2a\u5206\u7c7b\/\u6253\u6807\u6a21\u578b&#xff0c;\u4f46\u5acc\u4ece\u5934\u8bad\u592a\u6162\u592a\u96be\u7684 Python \u5de5\u7a0b\u5e08&#xff1b;\u4ee5\u53ca\u60f3\u641e\u6e05\u695a\u201c\u9884\u8bad\u7ec3\u6a21\u578b\u600e\u4e48\u4e3a\u6211\u6240\u7528\u201d\u7684\u4eba\u3002<\/p>\n<p>\u672c\u6587\u6982\u89c8&#xff1a;\u8fc1\u79fb\u5b66\u4e60&#xff08;Transfer Learning&#xff09;\u7684\u672c\u8d28\u662f\u201c\u7ad9\u5728\u9884\u8bad\u7ec3\u6a21\u578b\u7684\u80a9\u8180\u4e0a\u201d\u2014\u2014\u4e0d\u4ece\u5934\u8bad&#xff0c;\u76f4\u63a5\u590d\u7528\u5b83\u5728\u6d77\u91cf\u6570\u636e\u4e0a\u5df2\u7ecf\u5b66\u5230\u7684\u8bed\u8a00\u7279\u5f81\u3002\u8fd9\u7bc7\u7528 Hugging Face Transformers \u5e93\u628a\u8fd9\u4ef6\u4e8b\u8dd1\u901a&#xff1a;\u5148\u7528 pipeline \u770b\u9884\u8bad\u7ec3\u6a21\u578b\u5f00\u7bb1\u5373\u7528\u7684\u80fd\u529b&#xff0c;\u518d\u7528 AutoModelForSequenceClassification \u505a\u201c\u6362\u5934\u201d&#xff0c;\u63a5\u7740\u7528 Trainer \u5fae\u8c03&#xff0c;\u6700\u540e\u95ed\u73af\u9a8c\u8bc1\u3002\u5168\u7a0b\u4ee5\u4e00\u6bb5\u53ef\u8fd0\u884c\u4ee3\u7801\u4e3a\u4e3b\u7ebf\u9010\u6bb5\u62c6\u89e3&#xff0c;\u7ed3\u5c3e\u7ed9\u51fa\u6700\u5c0f\u53ef\u8fd0\u884c\u6a21\u677f\u548c\u5e38\u89c1\u6539\u52a8\u70b9\u3002<\/p>\n<hr \/>\n<h3>\u76ee\u5f55<\/h3>\n<ul>\n<li>\u4e00\u3001\u4ece\u96f6\u8bad\u7ec3\u4e3a\u4f55\u884c\u4e0d\u901a&#xff1f;\u5c0f\u6570\u636e\u96c6\u7684\u56f0\u5883<\/li>\n<li>\u4e8c\u3001\u8fc1\u79fb\u5b66\u4e60\u5230\u5e95\u5728\u201c\u8fc1\u79fb\u201d\u4ec0\u4e48&#xff1f;<\/li>\n<li>\u4e09\u3001pipeline \u4e09\u884c\u5c31\u80fd\u7528\u9884\u8bad\u7ec3\u6a21\u578b&#xff1f;<\/li>\n<li>\u56db\u3001AutoModel \u600e\u4e48\u201c\u6362\u5934\u201d&#xff1a;\u8fc1\u79fb\u5b66\u4e60\u7684\u6838\u5fc3\u52a8\u4f5c<\/li>\n<li>\u4e94\u3001\u51c6\u5907\u6570\u636e&#xff1a;Tokenizer \u4e0e Dataset<\/li>\n<li>\u516d\u3001Trainer \u5fae\u8c03&#xff1a;\u8ba9\u9884\u8bad\u7ec3\u6a21\u578b\u9002\u914d\u4f60\u7684\u4efb\u52a1<\/li>\n<li>\u4e03\u3001\u95ed\u73af\u9a8c\u8bc1&#xff1a;\u52a0\u8f7d\u5fae\u8c03\u540e\u7684\u6a21\u578b\u518d\u63a8\u7406<\/li>\n<li>\u516b\u3001Feature Extraction \u8fd8\u662f Fine-tuning&#xff1f;\u600e\u4e48\u9009<\/li>\n<li>\u4e5d\u3001\u5fae\u8c03\u597d\u7684\u6a21\u578b\u600e\u4e48\u8fdb\u5927\u6a21\u578b\u7ba1\u7ebf<\/li>\n<li>\u5341\u3001\u8fc1\u79fb\u5b66\u4e60\u7684\u8fb9\u754c&#xff1a;\u8fd9\u51e0\u79cd\u60c5\u51b5\u522b\u786c\u4e0a<\/li>\n<li>\u6700\u5c0f\u53ef\u8fd0\u884c\u6a21\u677f\u4e0e\u5e38\u89c1\u6539\u52a8\u70b9<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e00\u3001\u4ece\u96f6\u8bad\u7ec3\u4e3a\u4f55\u884c\u4e0d\u901a&#xff1f;\u5c0f\u6570\u636e\u96c6\u7684\u56f0\u5883<\/h3>\n<p>\u4f60\u63a5\u4e86\u4e2a\u5de5\u5355\u5206\u7c7b\u7684\u9700\u6c42&#xff1a;\u628a\u7528\u6237\u53cd\u9988\u5206\u6210 bug \/ \u54a8\u8be2 \/ \u5efa\u8bae \/ \u6295\u8bc9 \u56db\u7c7b\u3002\u624b\u91cc\u53ea\u6709 2000 \u6761\u6807\u6ce8\u3002<\/p>\n<p>\u5982\u679c\u4ece\u5934\u8bad\u4e00\u4e2a Transformer&#xff1a;<\/p>\n<ul>\n<li>\u968f\u673a\u521d\u59cb\u5316&#xff1a;\u6a21\u578b\u8981\u4ece\u201c\u4ec0\u4e48\u662f\u8bcd\u3001\u4ec0\u4e48\u662f\u53e5\u6cd5\u201d\u5f00\u59cb\u5b66&#xff0c;\u800c\u4f60\u53ea\u6709 2000 \u6761\u6837\u672c&#xff1b;<\/li>\n<li>\u5bb9\u6613\u8fc7\u62df\u5408&#xff1a;\u5c0f\u6570\u636e\u5582\u7ed9\u5927\u6a21\u578b&#xff0c;\u8bb0\u4f4f\u7684\u662f\u6837\u672c\u800c\u4e0d\u662f\u89c4\u5f8b&#xff1b;<\/li>\n<li>\u6162\u4e14\u4e0d\u7a33&#xff1a;\u5355\u5361\u8bad\u4e0a\u534a\u5929&#xff0c;\u51c6\u786e\u7387\u53ef\u80fd\u8fd8\u4e0d\u5982\u62cd\u8111\u888b\u89c4\u5219\u3002<\/li>\n<\/ul>\n<p>\u800c\u8fc1\u79fb\u5b66\u4e60\u7684\u601d\u8def\u662f&#xff1a;\u5df2\u7ecf\u6709\u4eba\u5728\u51e0\u4ebf\u6761\u6587\u672c\u4e0a\u8bad\u597d\u4e86\u4e00\u4e2a\u8bed\u8a00\u6a21\u578b&#xff0c;\u5b83\u65e9\u5c31\u61c2\u201c\u8bcd\u6027\u3001\u53e5\u6cd5\u3001\u8bed\u4e49\u201d\u4e86\u3002\u4f60\u53ea\u8981\u628a\u5b83\u5b66\u5230\u7684\u201c\u8bed\u8a00\u80fd\u529b\u201d\u501f\u8fc7\u6765&#xff0c;\u5957\u4e0a\u4f60\u81ea\u5df1\u7684\u5206\u7c7b\u5934\u5c31\u884c\u3002&#xff08;\u8fd9\u5957\u601d\u8def\u5728 CV \u91cc\u540c\u6837\u6210\u7acb&#xff1a;ResNet \u5728 ImageNet \u4e0a\u9884\u8bad\u7ec3\u597d\u540e&#xff0c;\u4f60\u6362\u6389\u6700\u540e\u7684\u5168\u8fde\u63a5\u5c42\u5c31\u80fd\u505a\u81ea\u5df1\u7684\u56fe\u50cf\u5206\u7c7b&#xff0c;\u5377\u79ef\u7279\u5f81\u76f4\u63a5\u590d\u7528\u2014\u2014NLP \u548c CV \u7684\u8fc1\u79fb\u5b66\u4e60\u662f\u540c\u4e00\u4e2a\u601d\u60f3\u7684\u4e24\u5957\u5b9e\u73b0\u3002&#xff09;<\/p>\n<p>\u5728\u5927\u6a21\u578b\u65f6\u4ee3&#xff0c;\u8fc1\u79fb\u5b66\u4e60\u4e0d\u4f46\u6ca1\u88ab\u53d6\u4ee3&#xff0c;\u53cd\u800c\u66f4\u5e95\u5c42\u4e86&#xff1a;\u4eca\u5929\u6240\u6709\u201c\u5fae\u8c03\u4e00\u4e2a\u57fa\u5ea7\u6a21\u578b\u201d\u7684\u505a\u6cd5&#xff08;LoRA\u3001\u6307\u4ee4\u5fae\u8c03\u3001\u9886\u57df\u9002\u914d&#xff09;\u672c\u8d28\u90fd\u662f\u8fc1\u79fb\u5b66\u4e60&#xff0c;\u53ea\u662f\u5934\u548c\u8bad\u7ec3\u65b9\u5f0f\u66f4\u82b1\u3002\u7406\u89e3\u5b83&#xff0c;\u662f\u770b\u61c2\u540e\u9762\u6574\u5957\u5fae\u8c03\u6280\u672f\u7684\u524d\u63d0\u2014\u2014\u8fd9\u4e5f\u662f\u5b83\u503c\u5f97\u5355\u72ec\u5199\u4e00\u7bc7\u7684\u539f\u56e0\u3002\u82e5\u786c\u8981\u4ece\u96f6\u8bad\u540c\u4e00\u4e2a\u5de5\u5355\u5206\u7c7b\u5668&#xff0c;\u4f60\u5f97 import torch.nn as nn \u624b\u5199 nn.Module&#xff08;\u5d4c\u5165\u5c42 &#043; Transformer \u7f16\u7801\u5668 &#043; \u7ebf\u6027\u5934&#xff09;\u3001\u81ea\u5df1\u9009\u521d\u59cb\u5316\u3001\u5199\u53cd\u5411\u4f20\u64ad\u3001\u8c03\u5b66\u4e60\u7387\u4e0e warmup\u2014\u2014\u51e0\u5341\u884c\u6837\u677f\u4ee3\u7801\u3001\u51e0\u5c0f\u65f6\u8bd5\u9519\u3002\u8fc1\u79fb\u5b66\u4e60\u628a\u8fd9\u4e9b\u5c01\u88c5\u8fdb from_pretrained&#xff0c;\u4f60\u53ea\u58f0\u660e num_labels&#061;4\u3002\u7701\u4e0b\u7684\u4e0d\u662f\u51e0\u884c\u4ee3\u7801&#xff0c;\u800c\u662f\u51e0\u4ebf\u6761\u6587\u672c\u7684\u8bad\u7ec3\u548c\u51e0\u5341\u6b21\u8bd5\u9519\u3002<\/p>\n<p>\u8fd9\u91cc\u6709\u4e2a\u5bb9\u6613\u5ffd\u7565\u7684\u70b9&#xff1a;\u4ece\u96f6\u8bad\u6162&#xff0c;\u4e0d\u53ea\u662f\u6162\u5728\u7b97\u529b&#xff0c;\u66f4\u6162\u5728\u201c\u8981\u8d70\u5b8c\u6574\u4e2a\u8bd5\u9519\u5468\u671f\u201d\u2014\u2014\u5b66\u4e60\u7387\u600e\u4e48\u8bbe\u3001\u8981\u4e0d\u8981 warmup\u3001\u6743\u91cd\u521d\u59cb\u5316\u7528\u54ea\u79cd\u3001\u8fc7\u62df\u5408\u4e86\u52a0\u4ec0\u4e48\u6b63\u5219&#xff0c;\u6bcf\u4e00\u9879\u90fd\u8981\u4f60\u81ea\u5df1\u8bd5\u3002\u8fc1\u79fb\u5b66\u4e60\u628a\u8fd9\u4e9b\u201c\u901a\u7528\u7ecf\u9a8c\u201d\u63d0\u524d\u5c01\u88c5\u8fdb\u9884\u8bad\u7ec3\u6743\u91cd&#xff0c;\u4f60\u5269\u4e0b\u7684\u65cb\u94ae\u53ea\u5269\u201c\u5b66\u4e60\u7387\u3001\u8f6e\u6570\u3001\u6279\u6b21\u201d\u4e09\u56db\u4e2a&#xff0c;\u8bd5\u9519\u6210\u672c\u9aa4\u964d\u3002<\/p>\n<p>\u56fe\u4e00&#xff1a;\u4ece\u96f6\u8bad\u7ec3 vs \u8fc1\u79fb\u5b66\u4e60<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260905043346-6a9b9baa9f75b.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u56fe\u4e00&#xff1a;\u5de6\u8fb9\u968f\u673a\u521d\u59cb\u5316 &#043; \u5c0f\u6570\u636e &#061; \u6162\u4e14\u6613\u8fc7\u62df\u5408&#xff1b;\u53f3\u8fb9\u9884\u8bad\u7ec3 Backbone \u590d\u7528 &#043; \u53ea\u8bad\u65b0\u5934 &#061; \u5feb\u4e14\u7a33&#xff09;<\/p>\n<h4>\u5c0f\u7ed3<\/h4>\n<p>\u8fc1\u79fb\u5b66\u4e60\u89e3\u51b3\u7684\u6838\u5fc3\u77db\u76fe\u662f&#xff1a;\u4f60\u7684\u6570\u636e\u5c11&#xff0c;\u4f46\u4efb\u52a1\u9700\u8981\u7684\u201c\u57fa\u7840\u8bed\u8a00\u80fd\u529b\u201d\u522b\u4eba\u5df2\u7ecf\u8bad\u597d\u4e86\u3002\u4e0b\u9762\u770b\u5b83\u5230\u5e95\u8fc1\u79fb\u4e86\u4ec0\u4e48\u3002<\/p>\n<hr \/>\n<h3>\u4e8c\u3001\u8fc1\u79fb\u5b66\u4e60\u5230\u5e95\u5728\u201c\u8fc1\u79fb\u201d\u4ec0\u4e48&#xff1f;<\/h3>\n<p>\u4e00\u4e2a\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u53ef\u4ee5\u62c6\u6210\u4e24\u534a&#xff1a;<\/p>\n<ul>\n<li>Backbone&#xff08;\u7279\u5f81\u63d0\u53d6\u5668&#xff09;&#xff1a;\u5e95\u5c42\u5230\u4e2d\u5c42\u5b66\u7684\u662f\u901a\u7528\u8bed\u8a00\u7279\u5f81\u2014\u2014\u8bcd\u6cd5\u3001\u53e5\u6cd5\u3001\u8bed\u4e49\u89d2\u8272\u3002\u8fd9\u90e8\u5206\u548c\u4f60\u5177\u4f53\u4efb\u52a1\u65e0\u5173&#xff0c;\u662f\u201c\u901a\u7528\u8d44\u4ea7\u201d\u3002<\/li>\n<li>Head&#xff08;\u4efb\u52a1\u5934&#xff09;&#xff1a;\u6700\u540e\u4e00\u5c42\u628a\u7279\u5f81\u6620\u5c04\u5230\u5177\u4f53\u8f93\u51fa&#xff08;\u6bd4\u5982 SST-2 \u7684\u6b63\u8d1f\u5411\u3001ImageNet \u7684 1000 \u7c7b&#xff09;\u3002\u8fd9\u90e8\u5206\u662f\u201c\u4efb\u52a1\u4e13\u5c5e\u201d\u7684\u3002\u4ece\u6570\u5b66\u770b&#xff0c;\u5206\u7c7b\u5934\u5c31\u662f\u4e00\u4e2a\u7ebf\u6027\u53d8\u6362 y &#061; W\u00b7h &#043; b&#xff1a;h \u662f Backbone \u8f93\u51fa\u7684\u53e5\u5411\u91cf&#xff0c;W\u3001b \u662f\u968f\u673a\u521d\u59cb\u5316\u7684\u53c2\u6570\u3002\u8bad\u7ec3\u65f6\u68af\u5ea6\u4ece y \u53cd\u4f20\u5230 W\u3001b&#xff0c;\u518d\u7ee7\u7eed\u5f80 Backbone \u4f20\u2014\u2014\u4f46\u56e0\u4e3a Backbone \u5b66\u4e60\u7387\u5c0f&#xff08;\u6216\u5e72\u8106\u51bb\u7ed3&#xff09;&#xff0c;\u5b83\u53ea\u88ab\u201c\u8f7b\u8f7b\u63a8\u201d&#xff0c;\u4e3b\u4f53\u7279\u5f81\u4e0d\u88ab\u7834\u574f\u3002\u7406\u89e3\u8fd9\u4e2a W\u00b7h&#043;b&#xff0c;\u5c31\u7406\u89e3\u4e86\u4e3a\u4ec0\u4e48\u201c\u6362\u5934 &#043; \u5c0f\u5b66\u4e60\u7387\u201d\u80fd\u6210\u7acb\u3002<\/li>\n<\/ul>\n<p>\u4e3a\u4ec0\u4e48 Backbone \u80fd\u76f4\u63a5\u590d\u7528&#xff1f;\u56e0\u4e3a\u9884\u8bad\u7ec3\u4efb\u52a1&#xff08;\u5982 MLM \u63a9\u7801\u8bed\u8a00\u5efa\u6a21\u3001NLI \u53e5\u5b50\u5173\u7cfb\u5224\u65ad&#xff09;\u903c\u6a21\u578b\u53bb\u7406\u89e3\u201c\u8bcd\u600e\u4e48\u7ec4\u6210\u53e5\u3001\u53e5\u600e\u4e48\u8868\u8fbe\u610f\u201d&#xff0c;\u5b66\u5230\u7684\u7279\u5f81\u662f\u4e0e\u5177\u4f53\u4e0b\u6e38\u4efb\u52a1\u89e3\u8026\u7684\u901a\u7528\u8bed\u8a00\u7ed3\u6784\u3002\u4f60\u505a\u5de5\u5355\u5206\u7c7b\u4e5f\u597d\u3001\u505a\u8206\u60c5\u5206\u6790\u4e5f\u597d&#xff0c;\u5e95\u5c42\u8981\u7406\u89e3\u7684\u201c\u8bed\u8a00\u201d\u662f\u540c\u4e00\u5957\u2014\u2014\u8fd9\u6b63\u662f\u8fc1\u79fb\u5b66\u4e60\u6210\u7acb\u7684\u524d\u63d0\u3002\u4ece\u65e9\u671f\u7684 Word2Vec&#xff08;\u53ea\u8bad\u9759\u6001\u8bcd\u5411\u91cf&#xff09;\u5230 ELMo&#xff08;\u53cc\u5411 LSTM \u51fa\u4e0a\u4e0b\u6587\u5411\u91cf&#xff09;\u3001\u518d\u5230 BERT&#xff08;Transformer &#043; MLM&#xff09;&#xff0c;\u9884\u8bad\u7ec3\u6a21\u578b\u4ece\u201c\u7ed9\u6bcf\u4e2a\u8bcd\u4e00\u4e2a\u5411\u91cf\u201d\u8fdb\u5316\u5230\u201c\u7ed9\u6bcf\u4e2a\u4e0a\u4e0b\u6587\u4e00\u4e2a\u8868\u793a\u201d&#xff0c;\u53ef\u590d\u7528\u7684\u7279\u5f81\u8d8a\u6765\u8d8a\u6df1&#xff0c;\u8fc1\u79fb\u5b66\u4e60\u7684\u6548\u679c\u4e5f\u8d8a\u6765\u8d8a\u597d\u2014\u2014\u4eca\u5929\u7528 from_pretrained \u4e00\u628a\u68ad&#xff0c;\u6b63\u662f\u8fd9\u6761\u6f14\u8fdb\u94fe\u7684\u7ec8\u70b9\u3002\u53cd\u8fc7\u6765&#xff0c;\u5982\u679c\u9884\u8bad\u7ec3\u8bed\u6599\u548c\u4f60\u7684\u4efb\u52a1\u57df\u5dee\u592a\u8fdc&#xff08;\u6bd4\u5982\u7528\u4ee3\u7801\u9884\u8bad\u7ec3\u6a21\u578b\u53bb\u505a\u533b\u7597\u6587\u672c\u5206\u7c7b&#xff09;&#xff0c;\u590d\u7528\u6536\u76ca\u5c31\u4f1a\u6253\u6298&#xff0c;\u8fd9\u70b9\u540e\u9762\u8fb9\u754c\u4e00\u8282\u4f1a\u8bb2\u3002<\/p>\n<p>\u8fc1\u79fb\u5b66\u4e60\u505a\u7684\u4e8b&#xff0c;\u5c31\u662f\u590d\u7528 Backbone&#xff0c;\u6362\u6389 Head&#xff1a;<\/p>\n<p>\u56fe\u4e8c&#xff1a;\u8fc1\u79fb\u5b66\u4e60\u7684\u7ed3\u6784<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260905043347-6a9b9bab6ac70.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u56fe\u4e8c&#xff1a;\u9884\u8bad\u7ec3 Backbone \u51bb\u7ed3\u4e0d\u52a8&#xff0c;\u65b0 Head \u968f\u673a\u521d\u59cb\u5316\u540e\u5355\u72ec\u8bad\u7ec3&#xff0c;\u68af\u5ea6\u53ea\u66f4\u65b0 Head&#xff09;<\/p>\n<p>\u5173\u952e\u70b9&#xff1a;\u65b0 Head \u662f\u968f\u673a\u521d\u59cb\u5316\u7684&#xff0c;\u9884\u8bad\u7ec3\u65f6\u6839\u672c\u4e0d\u5b58\u5728\u5b83&#xff0c;\u6240\u4ee5\u5fc5\u987b\u8bad\u7ec3\u3002\u800c Backbone \u5df2\u7ecf\u5f88\u61c2\u8bed\u8a00\u4e86&#xff0c;\u901a\u5e38\u51bb\u7ed3\u6216\u5c0f\u5e45\u5fae\u8c03\u5373\u53ef\u3002<\/p>\n<h4>\u5c0f\u7ed3<\/h4>\n<p>\u201c\u8fc1\u79fb\u201d\u8fc1\u79fb\u7684\u662f Backbone \u7684\u901a\u7528\u8bed\u8a00\u7279\u5f81&#xff1b;Head \u662f\u4f60\u81ea\u5df1\u7684\u4efb\u52a1\u5c42&#xff0c;\u5fc5\u987b\u91cd\u65b0\u8bad\u3002\u8fd9\u5c31\u662f\u540e\u9762\u6240\u6709\u4ee3\u7801\u7684\u5e95\u5c42\u903b\u8f91\u3002<\/p>\n<hr \/>\n<h3>\u4e09\u3001pipeline \u4e09\u884c\u5c31\u80fd\u7528\u9884\u8bad\u7ec3\u6a21\u578b&#xff1f;<\/h3>\n<p>\u5148\u4e0d\u6025\u7740\u5fae\u8c03&#xff0c;\u7528 pipeline \u770b\u4e00\u773c\u201c\u9884\u8bad\u7ec3\u6a21\u578b\u5230\u5e95\u6709\u591a\u80fd\u6253\u201d\u2014\u2014\u5b83\u751a\u81f3\u80fd\u5728\u6ca1\u89c1\u8fc7\u4f60\u6570\u636e\u7684\u60c5\u51b5\u4e0b\u505a\u96f6\u6837\u672c\u5206\u7c7b\u3002\u5f00\u59cb\u524d\u5148\u88c5\u4f9d\u8d56&#xff1a;pip install transformers datasets accelerate evaluate\u3002<\/p>\n<p><span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> pipeline<\/p>\n<p><span class=\"token comment\"># \u96f6\u6837\u672c\u5206\u7c7b&#xff1a;\u9884\u8bad\u7ec3 NLI \u6a21\u578b\u6ca1\u89c1\u8fc7\u201c\u5de5\u5355\u201d&#xff0c;\u5374\u56e0\u4e3a\u61c2\u8bed\u8a00\u5c31\u80fd\u5206<\/span><br \/>\ncls <span class=\"token operator\">&#061;<\/span> pipeline<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;zero-shot-classification&#034;<\/span><span class=\"token punctuation\">,<\/span> model<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;facebook\/bart-large-mnli&#034;<\/span><span class=\"token punctuation\">)<\/span><br \/>\nout <span class=\"token operator\">&#061;<\/span> cls<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;\u767b\u5f55\u540e\u9875\u9762\u4e00\u76f4\u8f6c\u5708&#034;<\/span><span class=\"token punctuation\">,<\/span> candidate_labels<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;bug&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;\u54a8\u8be2&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;\u5efa\u8bae&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span>out<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;labels&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> out<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;scores&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span>   <span class=\"token comment\"># \u2192 bug \/ 0.9x<\/span><\/p>\n<p>\u4e09\u884c\u5c31\u8dd1\u901a&#xff0c;\u8bf4\u660e\u4e00\u4ef6\u4e8b&#xff1a;\u9884\u8bad\u7ec3\u6a21\u578b\u5df2\u7ecf\u628a\u201c\u8bed\u8a00\u7406\u89e3\u201d\u8fd9\u9879\u57fa\u7840\u80fd\u529b\u8bad\u597d\u4e86&#xff0c;\u4f60\u770b\u5230\u7684\u201c\u5f00\u7bb1\u5373\u7528\u201d\u5c31\u662f\u4e0a\u6e38\u8fc1\u79fb\u5b66\u4e60\u7684\u7ed3\u679c\u3002<\/p>\n<p>pipeline \u672c\u8eab\u662f\u628a\u4e09\u4ef6\u4e8b\u5305\u6210\u4e86\u4e00\u6b65&#xff1a;\u7528\u5bf9\u5e94 AutoTokenizer \u5206\u8bcd \u2192 \u8c03 AutoModel \u62ff logits \u2192 \u7528\u540e\u5904\u7406\u628a logits \u53d8\u6210\u53ef\u8bfb\u6807\u7b7e\u548c\u5206\u6570\u3002\u5b83\u7701\u7684\u662f\u201c\u6837\u677f\u4ee3\u7801\u201d&#xff0c;\u5e95\u5c42\u8fd8\u662f Auto \u90a3\u4e00\u5957&#xff1b;\u7b49\u4f60\u8981\u628a\u63a7\u6bcf\u4e00\u6b65&#xff08;\u6bd4\u5982\u81ea\u5b9a\u4e49\u622a\u65ad\u3001\u52a0\u7279\u6b8a token\u3001\u6539\u8f93\u51fa\u683c\u5f0f&#xff09;\u65f6&#xff0c;\u5c31\u9000\u56de AutoTokenizer &#043; AutoModel \u624b\u52a8\u5199\u2014\u2014\u4e0b\u4e00\u8282\u6b63\u662f\u8fd9\u4e48\u505a\u7684\u3002<\/p>\n<p>\u9996\u6b21\u8fd0\u884c\u4f1a\u81ea\u52a8\u4e0b\u8f7d\u6a21\u578b\u6743\u91cd&#xff08;\u5982 bart-large-mnli \u7ea6 1.6GB&#xff09;&#xff0c;\u4e4b\u540e\u8d70 ~\/.cache\/huggingface \u672c\u5730\u7f13\u5b58&#xff0c;\u91cd\u590d\u8fd0\u884c\u4e0d\u91cd\u590d\u4e0b&#xff1b;\u79bb\u7ebf\u6216\u5185\u7f51\u73af\u5883&#xff0c;\u63d0\u524d\u7528 huggingface-cli download facebook\/bart-large-mnli \u628a\u6743\u91cd\u62c9\u8fdb\u7f13\u5b58\u5373\u53ef&#xff0c;\u907f\u514d\u8bad\u7ec3\u65f6\u5361\u5728\u7f51\u7edc\u3002<\/p>\n<p>pipeline \u9002\u5408\u201c\u5feb\u901f\u9a8c\u8bc1\u201d\u548c\u201c\u7b80\u5355\u4efb\u52a1\u201d\u3002\u4f46\u4f60\u7684\u5de5\u5355\u6709 4 \u7c7b\u3001\u6709\u81ea\u5df1\u7684\u6807\u6ce8\u5206\u5e03&#xff0c;\u96f6\u6837\u672c\u4e0d\u4e00\u5b9a\u51c6\u2014\u2014\u96f6\u6837\u672c\u9760\u7684\u662f\u201c\u8bed\u4e49\u5339\u914d\u201d&#xff0c;\u7c7b\u522b\u8fb9\u754c\u6a21\u7cca\u6216\u4f60\u7684\u6807\u7b7e\u4f53\u7cfb\u5f88\u4e13\u4e1a\u65f6&#xff0c;\u51c6\u786e\u7387\u4f1a\u6389\u3002\u8981\u771f\u6b63\u9002\u914d\u4f60\u7684\u6570\u636e&#xff0c;\u5f97\u81ea\u5df1\u505a\u201c\u6362\u5934 &#043; \u5fae\u8c03\u201d\u3002<\/p>\n<h4>\u5c0f\u7ed3<\/h4>\n<p>pipeline \u662f\u9884\u8bad\u7ec3\u80fd\u529b\u7684\u201c\u5feb\u6377\u5165\u53e3\u201d&#xff1b;\u96f6\u6837\u672c\u80fd\u6551\u6025\u4f46\u4e0a\u9650\u6709\u9650&#xff0c;\u8981\u9002\u914d\u81ea\u6709\u6570\u636e&#xff0c;\u4e0b\u4e00\u6b65\u7528 AutoModel \u663e\u5f0f\u6362\u5934\u3002<\/p>\n<hr \/>\n<h3>\u56db\u3001AutoModel \u600e\u4e48\u201c\u6362\u5934\u201d&#xff1a;\u8fc1\u79fb\u5b66\u4e60\u7684\u6838\u5fc3\u52a8\u4f5c<\/h3>\n<p>\u8fc1\u79fb\u5b66\u4e60\u7684\u6838\u5fc3\u52a8\u4f5c\u5c31\u4e00\u884c\u2014\u2014\u52a0\u8f7d\u9884\u8bad\u7ec3 Backbone&#xff0c;\u5e76\u6307\u5b9a\u4f60\u81ea\u5df1\u7684\u7c7b\u522b\u6570&#xff1a;<\/p>\n<p><span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> AutoModelForSequenceClassification<span class=\"token punctuation\">,<\/span> AutoTokenizer<\/p>\n<p>MODEL <span class=\"token operator\">&#061;<\/span> <span class=\"token string\">&#034;distilbert-base-uncased&#034;<\/span><br \/>\ntokenizer <span class=\"token operator\">&#061;<\/span> AutoTokenizer<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span>MODEL<span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># num_labels&#061;4&#xff1a;\u6211\u4eec\u7684\u5de5\u5355\u6709 4 \u7c7b\u3002<\/span><br \/>\n<span class=\"token comment\"># \u8fd9\u4e00\u5c42&#xff08;\u5206\u7c7b\u5934&#xff09;\u662f\u968f\u673a\u521d\u59cb\u5316\u7684\u201c\u65b0\u5934\u201d&#xff0c;\u9884\u8bad\u7ec3\u65f6\u4e0d\u5b58\u5728&#xff0c;\u6240\u4ee5\u5fc5\u987b\u8bad\u7ec3\u3002<\/span><br \/>\nmodel <span class=\"token operator\">&#061;<\/span> AutoModelForSequenceClassification<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span>MODEL<span class=\"token punctuation\">,<\/span> num_labels<span class=\"token operator\">&#061;<\/span><span class=\"token number\">4<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>AutoModelForSequenceClassification \u505a\u7684\u4e8b\u5f88\u660e\u786e&#xff1a;<\/p>\n<ul>\n<li>\u4e0b\u8f7d distilbert-base-uncased \u7684\u9884\u8bad\u7ec3\u6743\u91cd\u5f53 Backbone&#xff1b;<\/li>\n<li>\u5728\u5b83\u4e0a\u9762\u62fc\u4e00\u4e2a\u968f\u673a\u521d\u59cb\u5316\u7684\u7ebf\u6027\u5206\u7c7b\u5934&#xff0c;\u8f93\u51fa\u7ef4\u5ea6 &#061; num_labels&#xff1b;<\/li>\n<li>Backbone \u6743\u91cd\u6765\u81ea\u9884\u8bad\u7ec3&#xff0c;Head \u6743\u91cd\u968f\u673a\u3002<\/li>\n<\/ul>\n<p>\u987a\u5e26\u4e00\u63d0&#xff0c;num_labels \u51b3\u5b9a\u7684\u662f\u5206\u7c7b\u5934\u7684\u8f93\u51fa\u7ef4\u5ea6\u3002\u4e8c\u5206\u7c7b\u6709\u4e24\u79cd\u5e38\u89c1\u5199\u6cd5&#xff1a;\u8bbe num_labels&#061;1 \u914d BCE \u635f\u5931&#xff0c;\u6216\u8bbe num_labels&#061;2 \u8d70 softmax\u2014\u2014\u770b\u4f60\u7684\u8bc4\u4f30\u4e60\u60ef\u3002\u5982\u679c\u4e00\u6761\u5de5\u5355\u80fd\u540c\u65f6\u662f\u201cbug\u201d\u548c\u201c\u6295\u8bc9\u201d&#xff08;\u591a\u6807\u7b7e&#xff09;&#xff0c;\u5219\u8981\u6362 BCEWithLogitsLoss\u3001\u8ba9 Head \u8f93\u51fa N \u4e2a\u72ec\u7acb sigmoid&#xff0c;\u4f46 from_pretrained(num_labels&#061;N) \u8fd9\u4e00\u884c\u5199\u6cd5\u4e0d\u53d8\u3002\u672c\u6587\u805a\u7126\u5355\u6807\u7b7e\u591a\u7c7b&#xff0c;\u591a\u6807\u7b7e\u53ea\u662f\u635f\u5931\u51fd\u6570\u548c\u8f93\u51fa\u5c42\u7684\u5c0f\u6539\u52a8\u3002<\/p>\n<p>\u8fd9\u91cc\u7684 Auto \u524d\u7f00\u503c\u5f97\u8bb0\u4f4f&#xff1a;AutoModelForSequenceClassification \u53ea\u662f Auto \u5bb6\u65cf\u4e00\u5458&#xff0c;\u8fd8\u6709 AutoModel&#xff08;\u62ff\u9690\u85cf\u72b6\u6001&#xff09;\u3001AutoModelForTokenClassification&#xff08;\u547d\u540d\u5b9e\u4f53\u8bc6\u522b&#xff09;\u3001AutoModelForQuestionAnswering&#xff08;\u62bd\u53d6\u5f0f\u95ee\u7b54&#xff09;\u7b49\u3002\u540c\u4e00\u4e2a MODEL \u540d&#xff0c;\u6362\u4e2a\u4e0d\u540c\u7684 AutoModelForX \u5c31\u80fd\u9002\u914d\u4e0d\u540c\u4efb\u52a1\u2014\u2014Backbone \u590d\u7528\u3001\u53ea\u6362\u5934&#xff0c;\u8fd9\u6b63\u662f\u8fc1\u79fb\u5b66\u4e60\u88ab API \u5316\u7684\u4f53\u73b0&#xff1a;\u4f60\u4e0d\u5fc5\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u91cd\u65b0\u7406\u89e3\u6a21\u578b\u7ed3\u6784&#xff0c;\u6846\u67b6\u66ff\u4f60\u63a5\u597d\u5934\u3002<\/p>\n<p>\u5e38\u7528\u57fa\u5ea7\u600e\u4e48\u9009&#xff1a;\u5feb\u901f\u9a8c\u8bc1\u7528 distilbert-base-uncased&#xff08;\u5c0f\u3001\u5feb&#xff09;&#xff1b;\u8981\u7cbe\u5ea6\u6362 bert-base-uncased&#xff1b;\u82f1\u6587\u66f4\u5f3a\u7528 roberta-base&#xff1b;\u4e2d\u6587\u7528 bert-base-chinese \u6216 hfl\/chinese-roberta-wwm-ext&#xff1b;\u957f\u6587\u672c&#xff08;&gt;512 token&#xff09;\u770b longformer \/ bigbird\u3002\u5173\u952e\u70b9&#xff1a;\u9009\u6a21\u578b\u53ea\u6539 MODEL \u8fd9\u4e00\u4e2a\u5b57\u7b26\u4e32&#xff0c;\u6362\u5934\u3001\u5206\u8bcd\u3001\u8bad\u7ec3\u4ee3\u7801\u5168\u4e0d\u53d8\u2014\u2014\u53c8\u662f\u201c\u6362\u5934\u4e0d\u6362\u6d41\u7a0b\u201d\u7684\u4f53\u73b0\u3002\u5efa\u8bae\u5148\u7528\u5c0f\u6a21\u578b\u628a\u6574\u6761\u94fe\u8def\u8dd1\u901a\u3001\u786e\u8ba4\u6807\u6ce8\u548c\u6307\u6807\u6ca1\u95ee\u9898&#xff0c;\u518d\u6362\u5927\u6a21\u578b\u63d0\u4e0a\u9650&#xff0c;\u907f\u514d\u4e00\u4e0a\u6765\u5c31\u548c\u5927\u6a21\u578b\u640f\u663e\u5b58\u3001\u51fa\u95ee\u9898\u8fd8\u96be\u5b9a\u4f4d\u3002<\/p>\n<p>\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u540e\u9762\u5fc5\u987b train()&#xff1a;Head \u662f\u778e\u7684&#xff0c;\u5f97\u9760\u4f60\u7684\u6570\u636e\u628a\u5b83\u8bad\u4eae\u3002\u800c Backbone \u5df2\u7ecf\u5f88\u61c2\u8bed\u8a00&#xff0c;\u901a\u5e38\u662f\u201c\u987a\u624b\u5fae\u8c03\u201d\u800c\u4e0d\u662f\u201c\u4ece\u5934\u5b66\u201d\u3002<\/p>\n<p>\u7ecf\u9a8c&#xff1a;AutoTokenizer \u5fc5\u987b\u548c AutoModel \u7528\u540c\u4e00\u4e2a\u6a21\u578b\u540d\u2014\u2014\u5206\u8bcd\u65b9\u5f0f\u8981\u548c\u9884\u8bad\u7ec3\u65f6\u4e00\u81f4&#xff0c;\u5426\u5219\u8f93\u5165\u9519\u4f4d&#xff0c;\u6a21\u578b\u76f4\u63a5\u778e\u3002<\/p>\n<p>\u60f3\u5148\u53ea\u8bad Head\u3001\u51bb\u7ed3 Backbone&#xff0c;\u628a\u4e0a\u9762\u90a3\u884c\u4e4b\u540e\u8865\u4e00\u53e5\u5373\u53ef&#xff1a;<\/p>\n<p><span class=\"token comment\"># model.base_model \u5c31\u662f\u9884\u8bad\u7ec3 Backbone&#xff1b;\u9501\u6b7b\u5b83&#xff0c;\u68af\u5ea6\u53ea\u66f4\u65b0\u5206\u7c7b\u5934<\/span><br \/>\n<span class=\"token keyword\">for<\/span> p <span class=\"token keyword\">in<\/span> model<span class=\"token punctuation\">.<\/span>base_model<span class=\"token punctuation\">.<\/span>parameters<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    p<span class=\"token punctuation\">.<\/span>requires_grad <span class=\"token operator\">&#061;<\/span> <span class=\"token boolean\">False<\/span><\/p>\n<p>requires_grad&#061;False \u544a\u8bc9 PyTorch\u201c\u8fd9\u5c42\u53c2\u6570\u4e0d\u53c2\u4e0e\u53cd\u5411\u4f20\u64ad\u3001\u4e0d\u66f4\u65b0\u201d\u2014\u2014\u8fd9\u5c31\u662f\u8fc1\u79fb\u5b66\u4e60\u91cc\u201c\u51bb\u7ed3\u201d\u7684\u5b9e\u73b0\u65b9\u5f0f\u3002\u540e\u7eed\u4e24\u9636\u6bb5\u5de5\u4f5c\u6d41\u7684\u7b2c\u4e00\u6b65\u5c31\u662f\u9760\u5b83\u5148\u8dd1\u901a baseline\u3002\u6ce8\u610f AutoModelForSequenceClassification \u7684 base_model \u624d\u662f Backbone&#xff0c;\u5206\u7c7b\u5934\u5728 model.classifier \/ model.score \u4e0a&#xff0c;\u522b\u51bb\u9519\u4e86\u5bf9\u8c61\u3002<\/p>\n<p>from_pretrained \u80cc\u540e\u5e72\u4e86\u4e09\u4ef6\u4e8b&#xff1a;\u4e0b\u8f7d\u914d\u7f6e\u6587\u4ef6&#xff08;\u6a21\u578b\u5c42\u6570\u3001\u9690\u85cf\u7ef4\u5ea6\u7b49&#xff09;\u3001\u4e0b\u8f7d\u6743\u91cd\u6587\u4ef6\u3001\u628a\u5b83\u4eec\u7f13\u5b58\u5230\u672c\u5730 ~\/.cache\/huggingface&#xff08;\u4e0b\u6b21\u4e0d\u91cd\u590d\u4e0b&#xff09;\u3002\u5bf9\u5e94\u7684 save_pretrained(&#034;.\/ticket-bert&#034;) \u5219\u628a\u914d\u7f6e &#043; \u6743\u91cd &#043; \u5206\u8bcd\u5668\u4e00\u8d77\u5199\u8fdb\u76ee\u5f55&#xff0c;\u6240\u4ee5\u7b2c\u4e03\u8282\u624d\u80fd\u7528 pipeline(model&#061;&#034;.\/ticket-bert&#034;) \u76f4\u63a5\u52a0\u8f7d\u2014\u2014\u8fd9\u4e24\u4e2a\u65b9\u6cd5\u662f\u201c\u4e0b\u8f7d\u201d\u548c\u201c\u843d\u5730\u201d\u7684\u4e00\u5bf9&#xff0c;\u8bb0\u4f4f\u5b83\u4fe9\u5c31\u8bb0\u4f4f\u4e86\u8fc1\u79fb\u5b66\u4e60\u7684\u8f93\u5165\u8f93\u51fa\u8fb9\u754c\u3002<\/p>\n<h4>\u5c0f\u7ed3<\/h4>\n<p>\u201c\u6362\u5934\u201d&#061; \u6307\u5b9a num_labels&#xff0c;\u8ba9\u9884\u8bad\u7ec3 Backbone \u63a5\u4e0a\u4f60\u7684\u4efb\u52a1\u5c42&#xff1b;Head \u968f\u673a\u3001\u5fc5\u987b\u8bad&#xff0c;\u8fd9\u662f\u8fc1\u79fb\u5b66\u4e60\u843d\u5730\u7684\u603b\u5f00\u5173\u3002requires_grad&#061;False \u5219\u8d1f\u8d23\u628a Backbone \u51bb\u4f4f\u3002<\/p>\n<hr \/>\n<h3>\u4e94\u3001\u51c6\u5907\u6570\u636e&#xff1a;Tokenizer \u4e0e Dataset<\/h3>\n<p>\u6a21\u578b\u53ea\u5403\u6570\u5b57\u5f20\u91cf&#xff0c;\u6587\u672c\u5f97\u5148\u5206\u8bcd\u3002\u7528 datasets \u5e93\u628a CSV \u5de5\u5355\u53d8\u6210\u6a21\u578b\u80fd\u5403\u7684\u683c\u5f0f&#xff1a;<\/p>\n<p><span class=\"token keyword\">from<\/span> datasets <span class=\"token keyword\">import<\/span> load_dataset<br \/>\n<span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> DataCollatorWithPadding<\/p>\n<p>ds <span class=\"token operator\">&#061;<\/span> load_dataset<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;csv&#034;<\/span><span class=\"token punctuation\">,<\/span> data_files<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">{<\/span><span class=\"token string\">&#034;train&#034;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;tickets_train.csv&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;test&#034;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;tickets_test.csv&#034;<\/span><span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token keyword\">def<\/span> <span class=\"token function\">tokenize<\/span><span class=\"token punctuation\">(<\/span>batch<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token comment\"># truncation&#xff1a;\u8d85\u957f\u622a\u65ad&#xff1b;max_length \u6309\u4f60\u7684\u6587\u672c\u957f\u5ea6\u5b9a<\/span><br \/>\n    <span class=\"token keyword\">return<\/span> tokenizer<span class=\"token punctuation\">(<\/span>batch<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;text&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> truncation<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span> max_length<span class=\"token operator\">&#061;<\/span><span class=\"token number\">128<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>tok <span class=\"token operator\">&#061;<\/span> ds<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">map<\/span><span class=\"token punctuation\">(<\/span>tokenize<span class=\"token punctuation\">,<\/span> batched<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span> remove_columns<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;text&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><br \/>\ntok <span class=\"token operator\">&#061;<\/span> tok<span class=\"token punctuation\">.<\/span>rename_column<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;label&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;labels&#034;<\/span><span class=\"token punctuation\">)<\/span>          <span class=\"token comment\"># Trainer \u8ba4 labels \u8fd9\u4e2a\u5217\u540d<\/span><br \/>\ncollator <span class=\"token operator\">&#061;<\/span> DataCollatorWithPadding<span class=\"token punctuation\">(<\/span>tokenizer<span class=\"token operator\">&#061;<\/span>tokenizer<span class=\"token punctuation\">)<\/span>  <span class=\"token comment\"># \u6309 batch \u52a8\u6001\u8865\u9f50&#xff0c;\u7701\u663e\u5b58<\/span><\/p>\n<p>\u51e0\u4e2a\u5bb9\u6613\u8e29\u7684\u70b9&#xff1a;<\/p>\n<ul>\n<li>truncation&#061;True \u5fc5\u52a0&#xff1a;\u4e0d\u52a0\u5feb\u957f\u6587\u672c\u4f1a\u8d85\u6a21\u578b\u6700\u5927\u957f\u5ea6\u76f4\u63a5\u62a5\u9519\u3002<\/li>\n<li>DataCollatorWithPadding \u800c\u975e\u5199\u6b7b padding&#xff1a;\u5b83\u6309\u6bcf\u4e2a batch \u91cc\u6700\u957f\u5e8f\u5217\u52a8\u6001\u8865\u9f50&#xff0c;\u6bd4\u5168\u91cf\u8865\u5230\u6700\u5927\u957f\u5ea6\u7701\u663e\u5b58\u3002<\/li>\n<li>\u5217\u540d labels&#xff1a;Trainer \u9ed8\u8ba4\u8bfb labels&#xff0c;\u4f60\u7684 CSV \u82e5\u53eb label \u8981 rename_column\u3002<\/li>\n<\/ul>\n<p>\u4f60\u7684 tickets_train.csv \u957f\u8fd9\u6837\u5c31\u884c&#xff0c;\u4e24\u5217\u8db3\u591f&#xff1a;<\/p>\n<p>text,label<br \/>\n&#034;\u767b\u5f55\u540e\u9875\u9762\u4e00\u76f4\u8f6c\u5708&#xff0c;\u5237\u65b0\u4e5f\u6ca1\u7528&#034;,bug<br \/>\n&#034;\u60f3\u95ee\u4e0b\u4e0a\u4e2a\u6708\u7684\u8d26\u5355\u600e\u4e48\u5bfc\u51fa&#034;,\u54a8\u8be2<br \/>\n&#034;\u5efa\u8bae\u52a0\u4e2a\u6df1\u8272\u6a21\u5f0f&#xff0c;\u665a\u4e0a\u592a\u523a\u773c&#034;,\u5efa\u8bae<br \/>\n&#034;\u4ed8\u4e86\u94b1\u4f1a\u5458\u6ca1\u5230\u8d26&#xff0c;\u8981\u6c42\u9000\u6b3e&#034;,\u6295\u8bc9<\/p>\n<p>label \u7528\u6574\u6570&#xff08;0\/1\/2\/3&#xff09;\u6700\u7701\u4e8b&#xff0c;\u548c num_labels&#061;4 \u4e00\u4e00\u5bf9\u5e94&#xff1b;\u7528\u6587\u5b57\u6807\u7b7e\u4e5f\u884c&#xff0c;Trainer \u4f1a\u6309\u51fa\u73b0\u987a\u5e8f\u7f16\u7801&#xff0c;\u4f46\u6574\u6570\u66f4\u53ef\u63a7\u3002\u5207\u5206\u6bd4\u4f8b\u4e0a&#xff0c;\u6570\u636e\u5c11\u5c31 8:2&#xff08;train:test&#xff09;&#xff0c;\u6570\u636e\u8fc7\u4e07\u53ef\u4ee5 9:1 \u751a\u81f3 95:5\u2014\u2014\u6d4b\u8bd5\u96c6\u53ea\u8981\u80fd\u7a33\u5b9a\u53cd\u6620\u51c6\u786e\u7387\u5c31\u884c&#xff0c;\u4e0d\u5fc5\u592a\u5927\u3002<\/p>\n<p>max_length \u600e\u4e48\u5b9a&#xff1f; \u522b\u62cd\u8111\u888b\u586b 128\u3002\u5148 tok.filter(lambda x: len(x[&#034;text&#034;]) &gt; 128) \u770b\u770b\u4f60\u7684\u6587\u672c\u957f\u5ea6\u5206\u5e03&#xff1a;\u82e5 95% \u7684\u53e5\u5b50\u90fd\u77ed\u4e8e 128&#xff0c;\u586b 128 \u5c31\u591f&#xff0c;\u66f4\u957f\u7684\u622a\u65ad\u4e0d\u4e8f&#xff1b;\u82e5\u5927\u91cf\u6837\u672c\u8d85 256&#xff0c;\u586b 128 \u4f1a\u780d\u6389\u5173\u952e\u4fe1\u606f\u3001\u4f24\u51c6\u786e\u7387\u3002\u539f\u5219\u662f\u201c\u8986\u76d6\u7edd\u5927\u591a\u6570\u6837\u672c\u7684\u6700\u5c0f\u957f\u5ea6\u201d\u2014\u2014\u77ed\u5219\u6d6a\u8d39\u3001\u957f\u5219\u6491\u663e\u5b58\u3002tokenizer \u8fd8\u4f1a\u81ea\u52a8\u8865 [CLS]\/[SEP] \u8fd9\u7c7b\u7279\u6b8a token&#xff08;BERT \u9760\u5b83\u4eec\u5224\u65ad\u53e5\u5b50\u8fb9\u754c&#xff09;&#xff0c;\u4f60\u4e0d\u7528\u624b\u52a8\u62fc&#xff0c;\u4f46\u8981\u77e5\u9053\u5b83\u4eec\u5728&#xff0c;\u957f\u5ea6\u8981\u7559 2 \u4e2a\u4f4d\u7f6e\u7684\u4f59\u91cf\u3002<\/p>\n<p>\u7c7b\u522b\u4e0d\u5e73\u8861&#xff1a;\u82e5\u201c\u6295\u8bc9\u201d\u53ea\u6709\u51e0\u5341\u6761\u800c\u5176\u4ed6\u7c7b\u4e0a\u5343&#xff0c;\u76f4\u63a5\u8bad\u6a21\u578b\u4f1a\u5012\u5411\u591a\u6570\u7c7b\u3002\u8865\u6551\u5728\u6570\u636e\u4fa7\u2014\u2014\u91c7\u6837\u4e0a\u5bf9\u5c11\u6570\u7c7b resample \u6216\u7528 WeightedRandomSampler&#xff0c;\u635f\u5931\u4e0a\u7ed9\u5c11\u6570\u7c7b\u66f4\u5927\u7684 class_weight\u2014\u2014\u8bc4\u4f30\u65f6\u770b macro-F1 \u800c\u975e accuracy\u3002\u8fd9\u4e0d\u6539\u53d8\u8fc1\u79fb\u5b66\u4e60\u672c\u8eab&#xff0c;\u662f\u8bad\u7ec3\u524d\u7684\u5e38\u89c4\u5904\u7406&#xff0c;\u548c\u6362\u5934\u3001\u5fae\u8c03\u662f\u6b63\u4ea4\u7684\u4e24\u4ef6\u4e8b\u3002<\/p>\n<h4>\u5c0f\u7ed3<\/h4>\n<p>\u6570\u636e\u4fa7\u7684\u56fa\u5b9a\u52a8\u4f5c&#xff1a;\u5206\u8bcd &#043; \u622a\u65ad &#043; \u6539\u540d labels &#043; \u52a8\u6001\u8865\u9f50&#xff0c;\u5916\u52a0\u6309\u771f\u5b9e\u957f\u5ea6\u5206\u5e03\u5b9a max_length\u3002\u51c6\u5907\u597d\u5c31\u80fd\u4ea4\u7ed9 Trainer\u3002<\/p>\n<hr \/>\n<h3>\u516d\u3001Trainer \u5fae\u8c03&#xff1a;\u8ba9\u9884\u8bad\u7ec3\u6a21\u578b\u9002\u914d\u4f60\u7684\u4efb\u52a1<\/h3>\n<p>Trainer \u628a\u8bad\u7ec3\u5faa\u73af&#xff08;\u524d\u5411\u3001\u53cd\u5411\u3001\u4f18\u5316\u3001\u8bc4\u4f30&#xff09;\u5168\u5305\u4e86&#xff0c;\u4f60\u53ea\u8981\u7ed9\u6a21\u578b\u3001\u6570\u636e\u548c\u914d\u7f6e&#xff1a;<\/p>\n<p><span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> Trainer<span class=\"token punctuation\">,<\/span> TrainingArguments<\/p>\n<p>args <span class=\"token operator\">&#061;<\/span> TrainingArguments<span class=\"token punctuation\">(<\/span><br \/>\n    output_dir<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;.\/ticket-bert&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    per_device_train_batch_size<span class=\"token operator\">&#061;<\/span><span class=\"token number\">16<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    num_train_epochs<span class=\"token operator\">&#061;<\/span><span class=\"token number\">3<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    learning_rate<span class=\"token operator\">&#061;<\/span><span class=\"token number\">2e-5<\/span><span class=\"token punctuation\">,<\/span>        <span class=\"token comment\"># \u6bd4\u4ece\u5934\u8bad\u5c0f 10~100 \u500d&#xff1a;\u53ea\u5fae\u8c03&#xff0c;\u4e0d\u7834\u574f\u9884\u8bad\u7ec3\u7279\u5f81<\/span><br \/>\n    logging_steps<span class=\"token operator\">&#061;<\/span><span class=\"token number\">20<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    evaluation_strategy<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;epoch&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p>trainer <span class=\"token operator\">&#061;<\/span> Trainer<span class=\"token punctuation\">(<\/span><br \/>\n    model<span class=\"token operator\">&#061;<\/span>model<span class=\"token punctuation\">,<\/span><br \/>\n    args<span class=\"token operator\">&#061;<\/span>args<span class=\"token punctuation\">,<\/span><br \/>\n    train_dataset<span class=\"token operator\">&#061;<\/span>tok<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;train&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    eval_dataset<span class=\"token operator\">&#061;<\/span>tok<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;test&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    tokenizer<span class=\"token operator\">&#061;<\/span>tokenizer<span class=\"token punctuation\">,<\/span><br \/>\n    data_collator<span class=\"token operator\">&#061;<\/span>collator<span class=\"token punctuation\">,<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><br \/>\ntrainer<span class=\"token punctuation\">.<\/span>train<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><br \/>\ntrainer<span class=\"token punctuation\">.<\/span>save_model<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;.\/ticket-bert&#034;<\/span><span class=\"token punctuation\">)<\/span>   <span class=\"token comment\"># \u843d\u5730&#xff1a;\u6743\u91cd &#043; \u5206\u8bcd\u5668\u90fd\u5b58\u8fdb\u76ee\u5f55<\/span><\/p>\n<p>learning_rate&#061;2e-5 \u662f\u70b9\u775b\u4e4b\u7b14&#xff1a;\u5fae\u8c03\u8981\u7528\u6bd4\u4ece\u5934\u8bad\u5c0f\u5f97\u591a\u7684\u5b66\u4e60\u7387\u3002\u539f\u56e0\u5728\u4e0b\u4e00\u8282\u8bb2\u3002<\/p>\n<p>\u600e\u4e48\u5224\u65ad\u8bad\u7ec3\u5728\u6536\u655b\u3001\u6ca1\u8fc7\u62df\u5408&#xff1f; \u76ef evaluation_strategy&#061;&#034;epoch&#034; \u5410\u51fa\u7684 eval_loss&#xff1a;\u5b83\u5e94\u5f53\u968f\u8f6e\u6570\u4e0b\u964d\u5e76\u8d8b\u5e73&#xff1b;\u5982\u679c train_loss \u4e00\u8def\u964d\u4f46 eval_loss \u5f00\u59cb\u56de\u5347&#xff0c;\u5c31\u662f\u8fc7\u62df\u5408\u4fe1\u53f7\u2014\u2014\u6b64\u65f6\u8981\u4e48\u5c11\u8bad\u4e00\u8f6e&#xff08;num_train_epochs \u51cf 1&#xff09;&#xff0c;\u8981\u4e48\u52a0 weight_decay&#xff08;\u5982 0.01&#xff09;\u6216 warmup_steps\u3002\u522b\u7b49 train_loss \u8d34\u5230 0 \u624d\u505c&#xff0c;\u90a3\u5f80\u5f80\u662f\u80cc\u4e0b\u4e86\u8bad\u7ec3\u96c6\u3002\u51c6\u786e\u7387\u4e0d\u591f\u9ad8\u65f6&#xff0c;\u5148\u786e\u8ba4\u6d4b\u8bd5\u96c6\u6807\u7b7e\u6ca1\u6807\u9519&#xff0c;\u518d\u8003\u8651\u6362\u66f4\u5927\u7684\u57fa\u5ea7&#xff08;distilbert \u2192 bert-base&#xff09;\u6216\u89e3\u51bb\u66f4\u591a\u5c42\u5fae\u8c03&#xff0c;\u800c\u4e0d\u662f\u76f2\u76ee\u52a0\u8f6e\u6570\u3002<\/p>\n<p>\u5176\u4ed6\u5e38\u7528\u65cb\u94ae&#xff08;\u6309\u9700\u52a0\u8fdb TrainingArguments&#xff09;&#xff1a;<\/p>\n<ul>\n<li>warmup_steps&#061;100&#xff1a;\u8bad\u7ec3\u524d\u82e5\u5e72\u6b65\u5b66\u4e60\u7387\u4ece 0 \u7ebf\u6027\u5347\u5230\u76ee\u6807\u503c&#xff0c;\u907f\u514d\u5f00\u5934\u5927\u68af\u5ea6\u628a\u9884\u8bad\u7ec3\u7279\u5f81\u51b2\u574f&#xff0c;\u548c\u201c\u5c0f\u5b66\u4e60\u7387\u201d\u662f\u540c\u4e00\u4e2a\u4fdd\u62a4\u601d\u8def\u3002<\/li>\n<li>weight_decay&#061;0.01&#xff1a;L2 \u6b63\u5219&#xff0c;\u6291\u5236\u8fc7\u62df\u5408&#xff0c;\u5fae\u8c03\u65f6\u57fa\u672c\u5fc5\u5f00\u3002<\/li>\n<li>fp16&#061;True \/ bf16&#061;True&#xff1a;\u534a\u7cbe\u5ea6\u8bad\u7ec3&#xff0c;\u663e\u5b58\u780d\u534a\u3001\u901f\u5ea6\u7ffb\u500d&#xff0c;\u6709\u652f\u6301\u7684\u8bad\u7ec3\u5361\u5efa\u8bae\u5f00\u3002<\/li>\n<li>gradient_accumulation_steps&#061;2&#xff1a;\u7b49\u6548\u628a\u6279\u6b21\u7ffb\u500d\u4f46\u663e\u5b58\u4e0d\u6da8&#xff0c;\u5c0f\u663e\u5b58\u8dd1\u5927\u6279\u6b21\u7684\u5e38\u7528\u6280\u5de7\u3002<\/li>\n<li>load_best_model_at_end&#061;True &#043; metric_for_best_model&#061;&#034;eval_loss&#034;&#xff1a;\u6bcf\u8f6e\u8bc4\u4f30\u540e\u81ea\u52a8\u4fdd\u7559\u6700\u4f18\u6743\u91cd&#xff0c;\u7701\u5f97\u4f60\u81ea\u5df1\u6311 epoch\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u65cb\u94ae\u90fd\u4e0d\u6539\u53d8\u201c\u8fc1\u79fb\u5b66\u4e60\u201d\u7684\u672c\u8d28&#xff0c;\u53ea\u662f\u8ba9\u5fae\u8c03\u66f4\u7a33\u66f4\u5feb\u2014\u2014\u5148\u8dd1\u901a\u4e0a\u9762\u7684\u6700\u5c0f\u914d\u7f6e&#xff0c;\u518d\u6309\u663e\u5b58\u548c\u6548\u679c\u9010\u4e2a\u52a0\u3002<\/p>\n<h4>\u5c0f\u7ed3<\/h4>\n<p>Trainer \u4e00\u53e5\u8bdd\u542f\u52a8\u5fae\u8c03&#xff1b;save_model \u628a\u6743\u91cd\u843d\u76d8\u3002\u771f\u6b63\u8981\u8c03\u7684\u53ea\u6709\u5b66\u4e60\u7387\u3001\u8f6e\u6570\u3001\u6279\u6b21\u8fd9\u51e0\u4e2a\u65cb\u94ae&#xff0c;\u5224\u65ad\u597d\u574f\u770b eval_loss \u66f2\u7ebf&#xff0c;\u5176\u4f59\u65cb\u94ae\u6309\u9700\u53e0\u52a0\u3002<\/p>\n<hr \/>\n<h3>\u4e03\u3001\u95ed\u73af\u9a8c\u8bc1&#xff1a;\u52a0\u8f7d\u5fae\u8c03\u540e\u7684\u6a21\u578b\u518d\u63a8\u7406<\/h3>\n<p>\u5fae\u8c03\u5b8c&#xff0c;\u76f4\u63a5\u62ff\u4fdd\u5b58\u7684\u76ee\u5f55\u5f53\u6a21\u578b\u7528&#xff0c;\u9a8c\u8bc1\u5b83\u771f\u7684\u5b66\u4f1a\u4e86\u4f60\u7684\u5de5\u5355&#xff1a;<\/p>\n<p><span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> pipeline<\/p>\n<p><span class=\"token comment\"># \u76f4\u63a5\u5582\u76ee\u5f55&#xff0c;pipeline \u4f1a\u81ea\u52a8\u8bfb\u91cc\u9762\u7684\u6743\u91cd &#043; \u5206\u8bcd\u5668<\/span><br \/>\ncls <span class=\"token operator\">&#061;<\/span> pipeline<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;text-classification&#034;<\/span><span class=\"token punctuation\">,<\/span> model<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;.\/ticket-bert&#034;<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span>cls<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;\u652f\u4ed8\u63a5\u53e3\u8d85\u65f6&#xff0c;\u7528\u6237\u6295\u8bc9&#034;<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span>   <span class=\"token comment\"># \u2192 \u6295\u8bc9 \/ 0.97<\/span><\/p>\n<p>\u8fd9\u4e00\u6b65\u628a\u201c\u8fc1\u79fb\u5b66\u4e60\u201d\u95ed\u73af\u4e86&#xff1a;\u9884\u8bad\u7ec3 Backbone&#xff08;\u901a\u7528\u8bed\u8a00\u529b&#xff09;&#043; \u4f60\u7684 Head&#xff08;\u5de5\u5355 4 \u7c7b&#xff09;&#061; \u4e00\u4e2a\u4e13\u5c5e\u5206\u7c7b\u5668\u3002\u4ece pipeline \u521d\u4f53\u9a8c\u5230 Trainer \u5fae\u8c03\u518d\u5230\u8fd9\u91cc\u63a8\u7406&#xff0c;\u4e3b\u7ebf\u5c31\u662f\u8fd9\u4e00\u6761\u4ee3\u7801\u94fe\u3002<\/p>\n<p>\u4e24\u4e2a\u843d\u5730\u65f6\u7684\u5751\u987a\u624b\u63d0\u9192&#xff1a;\u5176\u4e00&#xff0c;\u5982\u679c\u4f60\u4e0d\u7528 pipeline\u3001\u800c\u662f\u76f4\u63a5 model(**inputs) \u62ff logits&#xff0c;\u63a8\u7406\u524d\u52a1\u5fc5 model.eval()\u2014\u2014\u5426\u5219 Dropout\/BatchNorm \u5904\u5728\u8bad\u7ec3\u6001&#xff0c;\u540c\u4e00\u6761\u6587\u672c\u591a\u6b21\u8dd1\u7ed3\u679c\u4f1a\u98d8\u3002\u5176\u4e8c&#xff0c;\u8bad\u597d\u7684\u6a21\u578b\u60f3\u5206\u4eab\u6216\u7248\u672c\u7ba1\u7406&#xff0c;\u7528 trainer.push_to_hub(&#034;\u4f60\u7684\u540d\/\u5de5\u5355\u6a21\u578b&#034;) \u4e00\u952e\u63a8\u5230 Hugging Face Hub&#xff0c;\u6bd4\u624b\u52a8\u4f20\u6587\u4ef6\u7701\u5fc3&#xff0c;\u56e2\u961f\u534f\u540c\u65f6\u5c24\u5176\u6709\u7528\u3002<\/p>\n<p>\u5176\u4e09&#xff0c;\u6279\u91cf\u63a8\u7406\u7528 tokenizer(texts, return_tensors&#061;&#034;pt&#034;, padding&#061;True, truncation&#061;True) \u4e00\u6b21\u5582\u591a\u6761&#xff0c;\u6bd4\u9010\u6761\u5faa\u73af\u5feb\u4e00\u4e2a\u6570\u91cf\u7ea7&#xff1b;padding&#061;True \u8ba9\u4e0d\u7b49\u957f\u6837\u672c\u5bf9\u9f50\u5230 batch \u5185\u6700\u957f&#xff0c;\u548c\u8bad\u7ec3\u65f6\u7684 DataCollatorWithPadding \u662f\u540c\u4e00\u4e2a\u8865\u9f50\u601d\u8def&#xff0c;\u53ea\u662f\u63a8\u7406\u65f6\u4f60\u624b\u52a8\u6307\u5b9a\u3002<\/p>\n<p>\u600e\u4e48\u5b9a\u91cf\u786e\u8ba4\u5b83\u771f\u5b66\u4f1a\u4e86&#xff1f; \u522b\u53ea\u770b\u4e00\u6761\u6837\u4f8b\u201c\u770b\u8d77\u6765\u5bf9\u201d&#xff0c;\u7528 evaluate \u5728\u6d4b\u8bd5\u96c6\u4e0a\u7b97\u51c6\u786e\u7387\u624d\u9760\u8c31&#xff1a;<\/p>\n<p><span class=\"token keyword\">import<\/span> evaluate<br \/>\nacc <span class=\"token operator\">&#061;<\/span> evaluate<span class=\"token punctuation\">.<\/span>load<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;accuracy&#034;<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token keyword\">def<\/span> <span class=\"token function\">compute_metrics<\/span><span class=\"token punctuation\">(<\/span>eval_pred<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    logits<span class=\"token punctuation\">,<\/span> labels <span class=\"token operator\">&#061;<\/span> eval_pred<br \/>\n    preds <span class=\"token operator\">&#061;<\/span> logits<span class=\"token punctuation\">.<\/span>argmax<span class=\"token punctuation\">(<\/span><span class=\"token operator\">&#8211;<\/span><span class=\"token number\">1<\/span><span class=\"token punctuation\">)<\/span>          <span class=\"token comment\"># \u53d6\u5206\u6570\u6700\u9ad8\u7684\u7c7b\u522b<\/span><br \/>\n    <span class=\"token keyword\">return<\/span> acc<span class=\"token punctuation\">.<\/span>compute<span class=\"token punctuation\">(<\/span>predictions<span class=\"token operator\">&#061;<\/span>preds<span class=\"token punctuation\">,<\/span> references<span class=\"token operator\">&#061;<\/span>labels<span class=\"token punctuation\">)<\/span><\/p>\n<p><span class=\"token comment\"># \u628a compute_metrics \u4f20\u7ed9 Trainer(&#8230;) &#xff0c;\u6bcf\u8f6e eval \u4f1a\u6253\u5370 accuracy<\/span><\/p>\n<p>\u7c7b\u522b\u4e0d\u5e73\u8861\u65f6\u628a accuracy \u6362\u6210 f1&#xff08;\u5b8f\u5e73\u5747&#xff09;&#xff0c;\u5426\u5219\u591a\u6570\u7c7b\u4f1a\u63a9\u76d6\u5c11\u6570\u7c7b\u7684\u5dee\u3002\u8fd9\u4e2a\u6307\u6807\u5c31\u662f\u4f60\u5224\u65ad\u201c\u8981\u4e0d\u8981\u89e3\u51bb\u5fae\u8c03\u3001\u5b66\u4e60\u7387\u8c03\u591a\u5c11\u201d\u7684\u5ba2\u89c2\u4f9d\u636e\u2014\u2014\u9760\u5b83\u800c\u4e0d\u662f\u9760\u611f\u89c9\u8c03\u53c2\u3002<\/p>\n<h4>\u5c0f\u7ed3<\/h4>\n<p>\u80fd\u7528\u81ea\u5df1\u7684\u76ee\u5f55\u8dd1\u51fa\u6b63\u786e\u7c7b\u522b&#xff0c;\u8bf4\u660e\u8fc1\u79fb\u5b66\u4e60\u843d\u5730\u6210\u529f\u3002\u76f4\u63a5\u8c03 model \u63a8\u7406\u8bb0\u5f97\u5148 model.eval()&#xff0c;\u8981\u5206\u4eab\u5c31 push_to_hub&#xff1b;\u5fae\u8c03\u51fa\u7684\u76ee\u5f55\u4e0b\u6e38\u7528 pipeline(&#034;text-classification&#034;, model&#061;&#8230;) \u5373\u53ef\u96f6\u6539\u52a8\u63a5\u5165&#xff0c;\u4e1a\u52a1\u65b9\u4e0d\u5fc5\u611f\u77e5\u80cc\u540e\u662f DistilBERT \u8fd8\u662f BERT\u3002\u7528 compute_metrics \u5728\u6d4b\u8bd5\u96c6\u4e0a\u91cf\u5316\u6548\u679c\u624d\u7a33\u3002\u4e0b\u9762\u770b\u4f60\u8be5\u7528\u54ea\u79cd\u5fae\u8c03\u7b56\u7565\u3002<\/p>\n<hr \/>\n<h3>\u516b\u3001Feature Extraction \u8fd8\u662f Fine-tuning&#xff1f;\u600e\u4e48\u9009<\/h3>\n<p>\u8fc1\u79fb\u5b66\u4e60\u843d\u5730\u6709\u4e24\u79cd\u529b\u5ea6&#xff0c;\u533a\u522b\u5728\u201cBackbone \u51bb\u4e0d\u51bb\u201d&#xff1a;<\/p>\n<p>\u56fe\u4e09&#xff1a;Feature Extraction vs Fine-tuning<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260905043348-6a9b9bac63a82.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u56fe\u4e09&#xff1a;\u5de6 &#061; \u51bb\u7ed3 Backbone \u53ea\u8bad Head&#xff1b;\u53f3 &#061; \u89e3\u51bb\u90e8\u5206\u5c42\u4e00\u8d77\u8bad&#xff0c;\u5b66\u4e60\u7387\u66f4\u5c0f&#xff09;<\/p>\n<table>\n<tr>\u7ef4\u5ea6Feature Extraction&#xff08;\u51bb\u7ed3&#xff09;Fine-tuning&#xff08;\u5fae\u8c03&#xff09;<\/tr>\n<tbody>\n<tr>\n<td>\u505a\u6cd5<\/td>\n<td>\u51bb\u7ed3 Backbone&#xff0c;\u53ea\u8bad Head<\/td>\n<td>\u89e3\u51bb\u90e8\u5206\/\u5168\u90e8\u5c42&#xff0c;\u5c0f\u5b66\u4e60\u7387\u4e00\u8d77\u8bad<\/td>\n<\/tr>\n<tr>\n<td>\u9002\u914d\u6570\u636e\u91cf<\/td>\n<td>\u5f88\u5c0f&#xff08;&lt; 1k&#xff09;<\/td>\n<td>\u4e2d<sub>\u5927&#xff08;1k<\/sub>100k&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u53ef\u8bad\u7ec3\u53c2\u6570<\/td>\n<td>\u7ea6 0.5%<\/td>\n<td>\u5168\u90e8<\/td>\n<\/tr>\n<tr>\n<td>\u5b66\u4e60\u7387<\/td>\n<td>\u8f83\u5927&#xff08;Head \u968f\u673a\u521d\u59cb\u5316&#xff09;<\/td>\n<td>\u5c0f&#xff08;1e-5 ~ 1e-4&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u98ce\u9669<\/td>\n<td>\u4e0a\u9650\u53d7 Backbone \u9650\u5236<\/td>\n<td>\u5b66\u4e60\u7387\u592a\u5927 \u2192 \u707e\u96be\u6027\u9057\u5fd8<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4e3a\u4ec0\u4e48\u5fae\u8c03\u8981\u7528\u66f4\u5c0f\u7684\u5b66\u4e60\u7387&#xff1f; \u9884\u8bad\u7ec3\u6743\u91cd\u5df2\u7ecf\u662f\u201c\u597d\u7279\u5f81\u201d&#xff0c;\u5982\u679c\u5b66\u4e60\u7387\u592a\u5927&#xff0c;\u51e0\u6b65\u5c31\u628a\u5b83\u4eec\u51b2\u574f&#xff0c;\u6a21\u578b\u5f97\u91cd\u65b0\u5b66\u8bed\u8a00\u2014\u2014\u8fd9\u53eb\u707e\u96be\u6027\u9057\u5fd8&#xff08;Catastrophic Forgetting&#xff09;\u3002\u5c0f\u5b66\u4e60\u7387\u8ba9 Backbone \u53ea\u505a\u201c\u5c0f\u5e45\u9002\u914d\u201d&#xff0c;\u4e0d\u7834\u574f\u5df2\u6709\u77e5\u8bc6\u3002<\/p>\n<p>\u7ecf\u9a8c\u6cd5\u5219&#xff1a;\u6570\u636e\u6781\u5c11\u5148\u51bb\u7ed3 Head \u8dd1\u901a&#xff1b;\u6570\u636e\u591f\u518d\u89e3\u51bb Backbone \u5fae\u8c03&#xff0c;\u5b66\u4e60\u7387\u4ece 2e-5 \u8d77\u8c03&#xff0c;\u770b eval_loss \u589e\u51cf\u3002\u6709\u516c\u5f00\u5b9e\u6d4b\u53ef\u53c2\u8003&#xff1a;\u5728 Oxford Flowers 102&#xff08;\u5c0f\u6570\u636e\u56fe\u50cf\u5206\u7c7b&#xff09;\u4e0a&#xff0c;\u51bb\u7ed3 Backbone \u7684\u201c\u7279\u5f81\u63d0\u53d6\u201d\u8fbe\u5230 85.7%&#xff0c;\u89e3\u51bb\u5fae\u8c03\u8fbe\u5230 92.5%\u2014\u2014\u90a3 7 \u4e2a\u70b9\u7684\u5dee\u8ddd&#xff0c;\u6b63\u6765\u81ea Backbone \u662f\u5426\u88ab\u5141\u8bb8\u9002\u914d\u4f60\u7684\u7279\u5b9a\u57df\u3002NLP \u4e0a\u540c\u7406&#xff1a;\u540c\u57df\u6570\u636e\u5fae\u8c03\u901a\u5e38\u6bd4\u7eaf\u51bb\u7ed3 Head \u9ad8\u51e0\u4e2a\u70b9&#xff0c;\u5dee\u8ddd\u5927\u5c0f\u53d6\u51b3\u4e8e\u4f60\u7684\u6570\u636e\u548c\u9884\u8bad\u7ec3\u57df\u7684\u8d34\u8fd1\u7a0b\u5ea6\u3002\u5230\u4e86\u5927\u6a21\u578b\u65f6\u4ee3&#xff0c;\u8fd9\u79cd\u201c\u89e3\u51bb\u90e8\u5206\u5c42\u5fae\u8c03\u201d\u88ab\u53d1\u626c\u6210 LoRA\u3001QLoRA \u7b49\u53c2\u6570\u9ad8\u6548\u5fae\u8c03&#xff08;PEFT&#xff09;&#xff1a;\u4e0d\u6539\u9884\u8bad\u7ec3\u6743\u91cd&#xff0c;\u53ea\u8bad\u4e00\u5c0f\u64ae\u4f4e\u79e9\u9002\u914d\u77e9\u9635&#xff0c;\u663e\u5b58\u4ece\u201c\u5168\u91cf\u5fae\u8c03\u201d\u964d\u5230\u51e0\u5206\u4e4b\u4e00\u2014\u2014\u601d\u60f3\u548c\u672c\u6587\u201c\u51bb\u7ed3 Backbone\u3001\u53ea\u8bad\u4e00\u5c0f\u5757\u201d\u4e00\u8109\u76f8\u627f&#xff0c;\u53ea\u662f\u9002\u914d\u7684\u201c\u5c0f\u5757\u201d\u66f4\u5de7\u3001\u66f4\u7701\u5361\u3002\u7406\u89e3\u8fc1\u79fb\u5b66\u4e60&#xff0c;\u5c31\u662f\u7406\u89e3\u8fd9\u5957 PEFT \u6280\u672f\u7684\u5730\u57fa\u3002<\/p>\n<p>\u63a8\u8350\u7684\u4e24\u9636\u6bb5\u5de5\u4f5c\u6d41&#xff08;\u7edd\u5927\u591a\u6570\u573a\u666f\u591f\u7528&#xff09;&#xff1a;<\/p>\n<li>\u9636\u6bb5\u4e00\u00b7\u51bb\u7ed3\u8dd1\u901a&#xff1a;\u5148 for p in model.base_model.parameters(): p.requires_grad &#061; False \u628a Backbone \u9501\u6b7b&#xff0c;\u53ea\u8bad Head&#xff0c;1~2 \u4e2a epoch\u3002\u8fd9\u4e00\u6b65\u51e0\u5206\u949f\u5c31\u80fd\u51fa baseline&#xff0c;\u7528\u6765\u9a8c\u8bc1\u201c\u6570\u636e\u6807\u6ce8\u3001\u5206\u8bcd\u3001\u6807\u7b7e\u6620\u5c04\u201d\u8fd9\u6761\u94fe\u8def\u672c\u8eab\u5bf9\u4e0d\u5bf9\u3002<\/li>\n<li>\u9636\u6bb5\u4e8c\u00b7\u89e3\u51bb\u5fae\u8c03&#xff1a;baseline \u8fbe\u6807\u540e&#xff0c;\u653e\u5f00 Backbone&#xff08;\u6216\u53ea\u653e\u6700\u540e\u51e0\u5c42&#xff09;&#xff0c;\u5b66\u4e60\u7387\u964d\u5230 2e-5 \u751a\u81f3 1e-5&#xff0c;\u518d\u8bad 2~3 \u4e2a epoch\u3002\u6b64\u65f6\u5728 baseline \u4e4b\u4e0a\u5c0f\u5e45\u63d0\u5347&#xff0c;\u4e14\u56e0\u4e3a\u5b66\u4e60\u7387\u5c0f&#xff0c;\u4e0d\u4f1a\u628a\u9884\u8bad\u7ec3\u7279\u5f81\u51b2\u574f\u3002<\/li>\n<p>\u4e24\u9636\u6bb5\u5206\u5f00\u505a\u7684\u597d\u5904&#xff1a;\u7b2c\u4e00\u9636\u6bb5\u80fd\u5feb\u901f\u66b4\u9732\u6570\u636e\/\u4ee3\u7801\u95ee\u9898&#xff0c;\u907f\u514d\u4f60\u82b1\u534a\u5c0f\u65f6\u5fae\u8c03\u5b8c\u624d\u53d1\u73b0\u6807\u7b7e\u5217\u540d\u5199\u9519\u3002<\/p>\n<h4>\u5c0f\u7ed3<\/h4>\n<p>\u9009\u54ea\u79cd\u770b\u6570\u636e\u91cf&#xff1a;\u5c11 \u2192 \u51bb\u7ed3 Head&#xff0c;\u591a \u2192 \u89e3\u51bb\u5fae\u8c03\u3002\u5fae\u8c03\u8bb0\u4f4f\u201c\u5c0f\u5b66\u4e60\u7387\u9632\u9057\u5fd8\u201d&#xff0c;\u5e76\u7528\u4e24\u9636\u6bb5\u5de5\u4f5c\u6d41\u5148\u9a8c\u8bc1\u94fe\u8def\u518d\u63d0\u7cbe\u5ea6\u3002<\/p>\n<hr \/>\n<h3>\u4e5d\u3001\u5fae\u8c03\u597d\u7684\u6a21\u578b\u600e\u4e48\u8fdb\u5927\u6a21\u578b\u7ba1\u7ebf<\/h3>\n<p>\u8fc1\u79fb\u5b66\u4e60\u4e0d\u662f\u201c\u88ab\u5927\u6a21\u578b\u53d6\u4ee3\u201d&#xff0c;\u800c\u662f\u5927\u6a21\u578b\u7ba1\u7ebf\u91cc\u7684\u9ad8\u6027\u4ef7\u6bd4\u96f6\u4ef6\u3002\u56db\u4e2a\u5177\u4f53\u843d\u70b9&#xff1a;<\/p>\n<p>\u573a\u666f 1&#xff1a;\u610f\u56fe\u8def\u7531\u2014\u2014\u7528\u5fae\u8c03\u5c0f\u6a21\u578b\u5224\u65ad\u7528\u6237 query \u610f\u56fe&#xff0c;\u51b3\u5b9a\u8fdb\u54ea\u4e2a Agent \/ \u5de5\u5177&#xff1a;<\/p>\n<p>router <span class=\"token operator\">&#061;<\/span> pipeline<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;text-classification&#034;<\/span><span class=\"token punctuation\">,<\/span> model<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;.\/ticket-bert&#034;<\/span><span class=\"token punctuation\">)<\/span><br \/>\nintent <span class=\"token operator\">&#061;<\/span> router<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;\u5e2e\u6211\u67e5\u4e00\u4e0b\u4e0a\u6708\u7684\u8d26\u5355&#034;<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;label&#034;<\/span><span class=\"token punctuation\">]<\/span>   <span class=\"token comment\"># \u2192 \u54a8\u8be2<\/span><br \/>\n<span class=\"token comment\"># if intent &#061;&#061; &#034;\u54a8\u8be2&#034;: call_billing_agent(user_text)<\/span><\/p>\n<p>\u573a\u666f 2&#xff1a;\u5927\u6a21\u578b\u515c\u5e95\u2014\u2014\u9ad8\u7f6e\u4fe1\u5ea6\u5c0f\u6a21\u578b\u76f4\u7b54&#xff0c;\u4f4e\u7f6e\u4fe1\u5ea6\u624d\u8fdb LLM&#xff0c;\u7701 token&#xff1a;<\/p>\n<p>out <span class=\"token operator\">&#061;<\/span> router<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;\u8fd9\u4e2a\u62a5\u9519\u600e\u4e48\u89e3&#034;<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><br \/>\n<span class=\"token keyword\">if<\/span> out<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;score&#034;<\/span><span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&gt;<\/span> <span class=\"token number\">0.9<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token keyword\">return<\/span> kb<span class=\"token punctuation\">[<\/span>out<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;label&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">]<\/span>      <span class=\"token comment\"># \u5c0f\u6a21\u578b\u76f4\u7b54&#xff0c;\u6beb\u79d2\u7ea7<\/span><br \/>\n<span class=\"token keyword\">else<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token keyword\">return<\/span> llm_chain<span class=\"token punctuation\">(<\/span>user_text<span class=\"token punctuation\">)<\/span>  <span class=\"token comment\"># \u4f4e\u7f6e\u4fe1\u5ea6\u624d\u8fdb\u5927\u6a21\u578b<\/span><\/p>\n<p>\u5b9e\u6d4b\u4e0a&#xff0c;\u5fae\u8c03\u5c0f\u6a21\u578b\u63a8\u7406 p99 \u5e38\u5728 10ms \u7ea7&#xff0c;\u800c\u4e00\u6b21 LLM \u8c03\u7528\u52a8\u8f84 1~2s\u2014\u2014\u628a 70% \u7684\u9ad8\u9891\u786e\u5b9a\u6027\u95ee\u9898\u62e6\u5728\u5feb\u8def\u5f84&#xff0c;\u6574\u4f53\u541e\u5410\u548c API \u6210\u672c\u90fd\u80fd\u964d\u4e00\u4e2a\u6570\u91cf\u7ea7&#xff0c;\u8fd9\u4e5f\u662f RAG \/ Agent \u7cfb\u7edf\u666e\u904d\u5728 LLM \u524d\u52a0\u4e00\u5c42\u5c0f\u6a21\u578b\u8fc7\u6ee4\u7684\u539f\u56e0\u3002<\/p>\n<p>\u573a\u666f 3&#xff1a;\u9886\u57df\u9002\u914d\u5668\u2014\u2014\u5728\u901a\u7528\u57fa\u5ea7\u4e0a\u5fae\u8c03\u51fa\u201c\u91d1\u878d \/ \u533b\u7597\u4e13\u7528\u201d\u5206\u7c7b\u5934&#xff0c;\u5f53\u5927\u6a21\u578b\u524d\u7f6e\u8fc7\u6ee4\u5668&#xff1a;\u5148\u7c97\u5206\u518d\u4ea4\u7ed9\u5bf9\u5e94\u4e13\u5bb6\u6a21\u578b\u3002\u6bd4\u5982\u533b\u7597\u95ee\u8bca\u5148\u5206\u51fa\u201c\u7528\u836f\u54a8\u8be2 \/ \u62a5\u544a\u89e3\u8bfb \/ \u6302\u53f7\u5bfc\u8bca\u201d&#xff0c;\u5927\u6a21\u578b\u53ea\u5904\u7406\u6700\u590d\u6742\u7684\u201c\u62a5\u544a\u89e3\u8bfb\u201d&#xff0c;\u5176\u4f59\u8d70\u8f7b\u91cf\u6d41\u7a0b&#xff0c;\u6574\u4f53\u65f6\u5ef6\u548c API \u6210\u672c\u90fd\u964d\u4e00\u622a\u3002<\/p>\n<p>\u573a\u666f 4&#xff1a;\u8bad\u7ec3\u6570\u636e\u521d\u6807\u2014\u2014\u7528\u5fae\u8c03\u6a21\u578b\u7ed9 LLM \u7684 SFT \u6570\u636e\u6253\u521d\u6807&#xff0c;\u4eba\u5de5\u53ea\u6821\u4e0d\u7528\u4ece\u96f6\u6807\u3002\u51e0\u5343\u6761\u5f85\u6807\u6ce8\u8bed\u6599&#xff0c;\u5c0f\u6a21\u578b\u5148\u6807\u4e00\u904d\u3001\u4eba\u5de5\u53ea\u6539\u9519\u7684&#xff0c;\u6bd4\u7eaf\u4eba\u5de5\u4ece\u767d\u7eb8\u5f00\u59cb\u5feb\u6570\u500d&#xff0c;\u6807\u6ce8\u6210\u672c\u964d 60%&#043;&#xff0c;\u4e14\u5148\u6709\u57fa\u7ebf\u6a21\u578b\u672c\u8eab\u5c31\u80fd\u53cd\u54fa\u6570\u636e\u8d28\u91cf\u68c0\u67e5&#xff08;\u6807\u9519\u7684\u548c\u6a21\u578b\u62ff\u4e0d\u51c6\u7684\u91cd\u5408\u5ea6\u6700\u9ad8&#xff09;\u3002<\/p>\n<p>\u56fe\u56db&#xff1a;\u5fae\u8c03\u6a21\u578b\u5728\u5927\u6a21\u578b\u7ba1\u7ebf\u4e2d\u7684\u4f4d\u7f6e<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260905043349-6a9b9bad2bac6.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>&#xff08;\u56fe\u56db&#xff1a;\u5fae\u8c03\u5c0f\u6a21\u578b\u5361\u5728\u8def\u7531 \/ \u515c\u5e95 \/ \u521d\u6807\u7b49\u73af\u8282&#xff0c;LLM \u53ea\u5904\u7406\u5b83\u641e\u4e0d\u5b9a\u7684\u590d\u6742\u6837\u672c&#xff09;<\/p>\n<h4>\u5c0f\u7ed3<\/h4>\n<p>\u5fae\u8c03\u5c0f\u6a21\u578b\u662f LLM \u7ba1\u7ebf\u7684\u201c\u5feb\u8def\u5f84\u201d&#xff1a;\u628a\u7b80\u5355\u3001\u9ad8\u9891\u3001\u4f4e\u7f6e\u4fe1\u5ea6\u95e8\u69db\u5185\u7684\u6d3b\u63fd\u4e86&#xff0c;\u590d\u6742\u6d3b\u7559\u7ed9\u5927\u6a21\u578b\u3002<\/p>\n<hr \/>\n<h3>\u5341\u3001\u8fc1\u79fb\u5b66\u4e60\u7684\u8fb9\u754c&#xff1a;\u8fd9\u51e0\u79cd\u60c5\u51b5\u522b\u786c\u4e0a<\/h3>\n<p>\u8fc1\u79fb\u5b66\u4e60\u4e0d\u662f\u4e07\u91d1\u6cb9\u3002\u4e09\u79cd\u573a\u666f\u4e0b\u5b83\u6536\u76ca\u6709\u9650&#xff0c;\u751a\u81f3\u4e0d\u5982\u4ece\u5934\u8bad&#xff1a;<\/p>\n<ul>\n<li>\u9884\u8bad\u7ec3\u57df\u548c\u4efb\u52a1\u57df\u5dee\u592a\u8fdc&#xff1a;\u7528\u4ee3\u7801\u8bed\u6599\u9884\u8bad\u7ec3\u7684\u6a21\u578b\u53bb\u505a\u53e4\u6587\u5206\u7c7b&#xff0c;Backbone \u7684\u901a\u7528\u8bed\u8a00\u7279\u5f81\u51e0\u4e4e\u7528\u4e0d\u4e0a\u3002\u57df\u5dee\u8ddd\u5927\u65f6&#xff0c;\u4f18\u5148\u9009\u540c\u57df\u9884\u8bad\u7ec3\u6a21\u578b&#xff08;\u4e2d\u6587\u4efb\u52a1\u7528 bert-base-chinese \u800c\u975e\u82f1\u6587 bert-base-uncased&#xff09;&#xff0c;\u5b9e\u5728\u6ca1\u6709\u518d\u8003\u8651\u4ece\u5934\u8bad\u6216\u505a\u9886\u57df\u7ee7\u7eed\u9884\u8bad\u7ec3\u3002<\/li>\n<li>\u4f60\u7684\u6570\u636e\u5176\u5b9e\u591f\u5927&#xff1a;\u5982\u679c\u5e72\u51c0\u6807\u6ce8\u6709\u51e0\u5341\u4e07\u6761&#xff0c;\u4ece\u96f6\u8bad\u4e00\u4e2a\u4e2d\u7b49\u6a21\u578b\u5f80\u5f80\u4e0a\u9650\u66f4\u9ad8\u2014\u2014\u8fc1\u79fb\u5b66\u4e60\u7684\u4f18\u52bf\u672c\u5c31\u662f\u201c\u7528\u5c0f\u6570\u636e\u8865\u77ed\u677f\u201d&#xff0c;\u6570\u636e\u7ba1\u591f\u65f6\u8fd9\u5757\u677f\u4e0d\u5b58\u5728\u3002<\/li>\n<li>\u4efb\u52a1\u672c\u8d28\u548c\u9884\u8bad\u7ec3\u76ee\u6807\u51b2\u7a81&#xff1a;\u9884\u8bad\u7ec3\u5b66\u7684\u662f\u201c\u7406\u89e3\u201d&#xff0c;\u5982\u679c\u4f60\u7684\u4efb\u52a1\u662f\u201c\u751f\u6210\u957f\u6587\u201d\u6216\u201c\u591a\u6a21\u6001\u5bf9\u9f50\u201d&#xff0c;\u5355\u7eaf\u6362\u5934\u5f0f\u8fc1\u79fb\u5b66\u4e60\u4e0d\u591f&#xff0c;\u5f97\u4e0a\u751f\u6210\u5f0f\u5fae\u8c03&#xff08;\u6307\u4ee4\u5fae\u8c03 \/ LoRA&#xff09;\u90a3\u5957&#xff0c;\u4e0d\u5728\u672c\u6587\u6362\u5934\u8303\u7574\u5185\u3002<\/li>\n<\/ul>\n<p>\u53e6\u6709\u4e00\u4e2a\u5e38\u89c1\u6df7\u6dc6&#xff1a;\u8fc1\u79fb\u5b66\u4e60 \u2260 \u6301\u7eed\u5b66\u4e60&#xff08;lifelong learning&#xff09;\u3002\u8fc1\u79fb\u5b66\u4e60\u662f\u201c\u4e00\u6b21\u9884\u8bad\u7ec3\u3001\u4e00\u6b21\u9002\u914d\u201d\u5c31\u5b9a\u683c&#xff0c;\u4ea7\u7269\u662f\u4e2a\u56fa\u5b9a\u7684\u5de5\u5355\u6a21\u578b&#xff1b;\u6301\u7eed\u5b66\u4e60\u5219\u8981\u6c42\u6a21\u578b\u5728\u4e0d\u65ad\u6765\u7684\u65b0\u4efb\u52a1\u4e0a\u4e00\u76f4\u5b66\u3001\u8fd8\u4e0d\u5fd8\u8bb0\u65e7\u4efb\u52a1&#xff0c;\u8981\u4e13\u95e8\u5bf9\u6297\u707e\u96be\u6027\u9057\u5fd8\u2014\u2014\u800c\u672c\u6587\u8bf4\u7684\u201c\u5c0f\u5b66\u4e60\u7387\u9632\u9057\u5fd8\u201d\u6070\u6070\u662f\u5728\u5355\u8f6e\u5fae\u8c03\u91cc\u907f\u514d\u628a\u9884\u8bad\u7ec3\u7279\u5f81\u51b2\u574f&#xff0c;\u548c\u6301\u7eed\u5b66\u4e60\u7684\u201c\u8de8\u4efb\u52a1\u4e0d\u5fd8\u201d\u662f\u4e24\u4ef6\u4e8b\u3002\u5b9e\u9645\u843d\u5730&#xff0c;\u65b0\u6765\u4e00\u6279\u6807\u6ce8\u76f4\u63a5\u91cd\u8bad\u6216\u589e\u91cf\u5fae\u8c03\u5373\u53ef&#xff0c;\u4e0d\u5fc5\u8ffd\u6c42\u201c\u4e00\u4e2a\u6a21\u578b\u6c38\u8fdc\u5b66\u201d&#xff0c;\u90a3\u4f1a\u5f15\u5165\u6301\u7eed\u5b66\u4e60\u7684\u989d\u5916\u590d\u6742\u5ea6&#xff0c;\u5bf9\u5c0f\u56e2\u961f\u5f80\u5f80\u5f97\u4e0d\u507f\u5931\u3002<\/p>\n<p>\u4e00\u53e5\u8bdd\u5224\u65ad&#xff1a;\u5c0f\u6570\u636e &#043; \u540c\u57df &#043; \u7406\u89e3\u7c7b\u4efb\u52a1 &#061; \u8fc1\u79fb\u5b66\u4e60\u6700\u8212\u670d\u7684\u533a\u95f4&#xff1b;\u8d85\u51fa\u8fd9\u4e2a\u533a\u95f4&#xff0c;\u5148\u60f3\u6e05\u695a\u8be5\u6362\u6a21\u578b\u3001\u52a0\u6570\u636e\u8fd8\u662f\u6362\u65b9\u6cd5&#xff0c;\u522b\u65e0\u8111 from_pretrained\u3002<\/p>\n<hr \/>\n<h3>\u6700\u5c0f\u53ef\u8fd0\u884c\u6a21\u677f\u4e0e\u5e38\u89c1\u6539\u52a8\u70b9<\/h3>\n<p>\u628a\u524d\u9762\u4e94\u6bb5\u62fc\u6210\u4e00\u6761\u6700\u5c0f\u53ef\u8fd0\u884c\u94fe&#xff08;\u4ee5\u5de5\u5355 4 \u5206\u7c7b\u4e3a\u4f8b&#xff09;&#xff1a;<\/p>\n<p><span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> <span class=\"token punctuation\">(<\/span>pipeline<span class=\"token punctuation\">,<\/span> AutoTokenizer<span class=\"token punctuation\">,<\/span><br \/>\n                          AutoModelForSequenceClassification<span class=\"token punctuation\">,<\/span><br \/>\n                          Trainer<span class=\"token punctuation\">,<\/span> TrainingArguments<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">from<\/span> datasets <span class=\"token keyword\">import<\/span> load_dataset<br \/>\n<span class=\"token keyword\">from<\/span> transformers <span class=\"token keyword\">import<\/span> DataCollatorWithPadding<\/p>\n<p>MODEL <span class=\"token operator\">&#061;<\/span> <span class=\"token string\">&#034;distilbert-base-uncased&#034;<\/span><br \/>\ntokenizer <span class=\"token operator\">&#061;<\/span> AutoTokenizer<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span>MODEL<span class=\"token punctuation\">)<\/span><br \/>\nmodel <span class=\"token operator\">&#061;<\/span> AutoModelForSequenceClassification<span class=\"token punctuation\">.<\/span>from_pretrained<span class=\"token punctuation\">(<\/span>MODEL<span class=\"token punctuation\">,<\/span> num_labels<span class=\"token operator\">&#061;<\/span><span class=\"token number\">4<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>ds <span class=\"token operator\">&#061;<\/span> load_dataset<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;csv&#034;<\/span><span class=\"token punctuation\">,<\/span> data_files<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">{<\/span><span class=\"token string\">&#034;train&#034;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;tickets_train.csv&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;test&#034;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;tickets_test.csv&#034;<\/span><span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">)<\/span><br \/>\ntok <span class=\"token operator\">&#061;<\/span> ds<span class=\"token punctuation\">.<\/span><span class=\"token builtin\">map<\/span><span class=\"token punctuation\">(<\/span><span class=\"token keyword\">lambda<\/span> b<span class=\"token punctuation\">:<\/span> tokenizer<span class=\"token punctuation\">(<\/span>b<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;text&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> truncation<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">,<\/span> max_length<span class=\"token operator\">&#061;<\/span><span class=\"token number\">128<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span><br \/>\n             batched<span class=\"token operator\">&#061;<\/span><span class=\"token boolean\">True<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">.<\/span>rename_column<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;label&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;labels&#034;<\/span><span class=\"token punctuation\">)<\/span><br \/>\ncollator <span class=\"token operator\">&#061;<\/span> DataCollatorWithPadding<span class=\"token punctuation\">(<\/span>tokenizer<span class=\"token operator\">&#061;<\/span>tokenizer<span class=\"token punctuation\">)<\/span><\/p>\n<p>trainer <span class=\"token operator\">&#061;<\/span> Trainer<span class=\"token punctuation\">(<\/span><br \/>\n    model<span class=\"token operator\">&#061;<\/span>model<span class=\"token punctuation\">,<\/span><br \/>\n    args<span class=\"token operator\">&#061;<\/span>TrainingArguments<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;.\/ticket-bert&#034;<\/span><span class=\"token punctuation\">,<\/span> per_device_train_batch_size<span class=\"token operator\">&#061;<\/span><span class=\"token number\">16<\/span><span class=\"token punctuation\">,<\/span><br \/>\n                           num_train_epochs<span class=\"token operator\">&#061;<\/span><span class=\"token number\">3<\/span><span class=\"token punctuation\">,<\/span> learning_rate<span class=\"token operator\">&#061;<\/span><span class=\"token number\">2e-5<\/span><span class=\"token punctuation\">,<\/span><br \/>\n                           evaluation_strategy<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;epoch&#034;<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    train_dataset<span class=\"token operator\">&#061;<\/span>tok<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;train&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> eval_dataset<span class=\"token operator\">&#061;<\/span>tok<span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;test&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    tokenizer<span class=\"token operator\">&#061;<\/span>tokenizer<span class=\"token punctuation\">,<\/span> data_collator<span class=\"token operator\">&#061;<\/span>collator<span class=\"token punctuation\">,<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><br \/>\ntrainer<span class=\"token punctuation\">.<\/span>train<span class=\"token punctuation\">(<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">;<\/span> trainer<span class=\"token punctuation\">.<\/span>save_model<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;.\/ticket-bert&#034;<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>cls <span class=\"token operator\">&#061;<\/span> pipeline<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;text-classification&#034;<\/span><span class=\"token punctuation\">,<\/span> model<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;.\/ticket-bert&#034;<\/span><span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span>cls<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;\u767b\u5f55\u540e\u767d\u5c4f&#034;<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>\u5e38\u89c1\u6539\u52a8\u70b9&#xff1a;<\/p>\n<ul>\n<li>\u6362\u6570\u636e\u96c6&#xff1a;CSV \u6539 JSON \/ \u76f4\u63a5 load_dataset(&#034;imdb&#034;) \u7b49\u5185\u7f6e\u96c6&#xff1b;\u5217\u540d\u4e0d\u540c\u5c31\u6539 rename_column\u3002<\/li>\n<li>\u6362\u6a21\u578b&#xff1a;distilbert-base-uncased \u2192 bert-base-chinese&#xff08;\u4e2d\u6587&#xff09;\u3001roberta-base&#xff08;\u82f1\u6587\u66f4\u5f3a&#xff09;&#xff1b;\u6362\u5b8c num_labels \u4e0d\u53d8\u3002<\/li>\n<li>\u8c03\u5b66\u4e60\u7387&#xff1a;2e-5 \u8d77&#xff1b;eval_loss \u4e0d\u964d\u5c31\u964d\u5230 1e-5&#xff0c;\u9707\u8361\u5c31\u52a0 warmup_steps\u3002<\/li>\n<li>\u52a0\u8bc4\u4f30\u6307\u6807&#xff1a;from evaluate import load; acc &#061; load(&#034;accuracy&#034;)&#xff0c;\u5728 compute_metrics \u91cc\u7b97 F1&#xff0c;\u7c7b\u522b\u4e0d\u5e73\u8861\u65f6\u522b\u53ea\u770b\u51c6\u786e\u7387\u3002<\/li>\n<li>\u591a\u6807\u7b7e\u5206\u7c7b&#xff1a;\u4e00\u6761\u6837\u672c\u53ef\u5c5e\u591a\u7c7b\u65f6&#xff0c;\u635f\u5931\u6362 BCEWithLogitsLoss&#xff0c;Head \u8f93\u51fa N \u4e2a\u72ec\u7acb sigmoid&#xff08;num_labels \u4e0d\u53d8&#xff09;&#xff0c;\u8bc4\u4f30\u7528 micro\/macro F1\u3002<\/li>\n<li>\u591a\u5361 \/ \u5927\u6279\u6b21&#xff1a;\u5355\u5361\u663e\u5b58\u4e0d\u591f\u5c31 gradient_accumulation_steps &#043; fp16 \u9876\u4e0a&#xff0c;\u6216 accelerate launch \u591a\u5361\u8d77\u8bad&#xff0c;TrainingArguments \u4e0d\u7528\u5927\u6539\u3002<\/li>\n<li>\u4e2d\u6587\u573a\u666f&#xff1a;\u7528 bert-base-chinese \u6216 hfl\/chinese-roberta-wwm-ext&#xff0c;\u5206\u8bcd\u5668\u548c\u6a21\u578b\u540d\u4fdd\u6301\u4e00\u81f4\u3002<\/li>\n<\/ul>\n<hr \/>\n<p>#\u8fc1\u79fb\u5b66\u4e60 #HuggingFace #Transformers #\u6a21\u578b\u5fae\u8c03 #\u9884\u8bad\u7ec3\u6a21\u578b<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730&#xff1f;Transformers \u5e93\u5fae\u8c03\u5b9e\u6218 \u5173\u952e\u8bcd&#xff1a;\u8fc1\u79fb\u5b66\u4e60\u3001Transfer Learning\u3001Hugging Face Transformers\u3001\u6a21\u578b\u5fae\u8c03&#xff08;fine-tuning&#xff09;\u3001\u9884\u8bad\u7ec3\u6a21\u578b \u9002\u8bfb\u4eba\u7fa4&#xff1a;\u624b\u91cc\u6709\u51e0\u767e\u5230\u51e0\u4e07\u6761\u6807\u6ce8\u6570\u636e\u3001\u60f3\u8bad\u4e2a\u5206\u7c7b\/\u6253\u6807\u6a21\u578b&#xff0c;\u4f46\u5acc\u4ece\u5934\u8bad\u592a\u6162\u592a\u96be\u7684 Python \u5de5\u7a0b\u5e08&#xff1b;\u4ee5\u53ca\u60f3\u641e\u6e05\u695a\u201c\u9884\u8bad\u7ec3\u6a21\u578b\u600e\u4e48\u4e3a\u6211\u6240\u7528\u201d\u7684\u4eba\u3002 \u672c\u6587\u6982\u89c8&amp;#xff1a<\/p>\n","protected":false},"author":2,"featured_media":100810,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[9109,841,50,207,224,2755],"topic":[],"class_list":["post-100814","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-9109","tag-transformer","tag-50","tag-207","tag-224","tag-2755"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730\uff1fTransformers \u5e93\u5fae\u8c03\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/100814.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730\uff1fTransformers \u5e93\u5fae\u8c03\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730&#xff1f;Transformers \u5e93\u5fae\u8c03\u5b9e\u6218 \u5173\u952e\u8bcd&#xff1a;\u8fc1\u79fb\u5b66\u4e60\u3001Transfer Learning\u3001Hugging Face Transformers\u3001\u6a21\u578b\u5fae\u8c03&#xff08;fine-tuning&#xff09;\u3001\u9884\u8bad\u7ec3\u6a21\u578b \u9002\u8bfb\u4eba\u7fa4&#xff1a;\u624b\u91cc\u6709\u51e0\u767e\u5230\u51e0\u4e07\u6761\u6807\u6ce8\u6570\u636e\u3001\u60f3\u8bad\u4e2a\u5206\u7c7b\/\u6253\u6807\u6a21\u578b&#xff0c;\u4f46\u5acc\u4ece\u5934\u8bad\u592a\u6162\u592a\u96be\u7684 Python \u5de5\u7a0b\u5e08&#xff1b;\u4ee5\u53ca\u60f3\u641e\u6e05\u695a\u201c\u9884\u8bad\u7ec3\u6a21\u578b\u600e\u4e48\u4e3a\u6211\u6240\u7528\u201d\u7684\u4eba\u3002 \u672c\u6587\u6982\u89c8&amp;#xff1a\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/100814.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-05T04:33:50+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260905043346-6a9b9baa9f75b.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"9 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/100814.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/100814.html\",\"name\":\"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730\uff1fTransformers \u5e93\u5fae\u8c03\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-05T04:33:50+00:00\",\"dateModified\":\"2026-09-05T04:33:50+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/100814.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/100814.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/100814.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730\uff1fTransformers \u5e93\u5fae\u8c03\u5b9e\u6218\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730\uff1fTransformers \u5e93\u5fae\u8c03\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/100814.html","og_locale":"zh_CN","og_type":"article","og_title":"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730\uff1fTransformers \u5e93\u5fae\u8c03\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730&#xff1f;Transformers \u5e93\u5fae\u8c03\u5b9e\u6218 \u5173\u952e\u8bcd&#xff1a;\u8fc1\u79fb\u5b66\u4e60\u3001Transfer Learning\u3001Hugging Face Transformers\u3001\u6a21\u578b\u5fae\u8c03&#xff08;fine-tuning&#xff09;\u3001\u9884\u8bad\u7ec3\u6a21\u578b \u9002\u8bfb\u4eba\u7fa4&#xff1a;\u624b\u91cc\u6709\u51e0\u767e\u5230\u51e0\u4e07\u6761\u6807\u6ce8\u6570\u636e\u3001\u60f3\u8bad\u4e2a\u5206\u7c7b\/\u6253\u6807\u6a21\u578b&#xff0c;\u4f46\u5acc\u4ece\u5934\u8bad\u592a\u6162\u592a\u96be\u7684 Python \u5de5\u7a0b\u5e08&#xff1b;\u4ee5\u53ca\u60f3\u641e\u6e05\u695a\u201c\u9884\u8bad\u7ec3\u6a21\u578b\u600e\u4e48\u4e3a\u6211\u6240\u7528\u201d\u7684\u4eba\u3002 \u672c\u6587\u6982\u89c8&amp;#xff1a","og_url":"https:\/\/www.wsisp.com\/helps\/100814.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-05T04:33:50+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260905043346-6a9b9baa9f75b.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"9 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/100814.html","url":"https:\/\/www.wsisp.com\/helps\/100814.html","name":"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730\uff1fTransformers \u5e93\u5fae\u8c03\u5b9e\u6218 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-05T04:33:50+00:00","dateModified":"2026-09-05T04:33:50+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/100814.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/100814.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/100814.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u8fc1\u79fb\u5b66\u4e60\u600e\u4e48\u843d\u5730\uff1fTransformers \u5e93\u5fae\u8c03\u5b9e\u6218"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/100814","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=100814"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/100814\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/100810"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=100814"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=100814"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=100814"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=100814"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}